Skip to main content

Tasks: Harness Engine

Canonical delivery phases (2026-10-21): See Delivery phases: Phase 1 independent engine/gateway (PR 2401); Phase 2 separate Deep Agents integration and pluggable builder; Phase 3 external sandbox execution; Phase 4 cross-harness composition. The numbered workstreams below preserve old task IDs and implementation history; they are not release phases or a reviewer approval record.

Input: Design documents in docs/docs/specs/2026-08-17-harness-engine/ Prerequisites: spec.md, plan.md, research.md, data-model.md, mongodb-migration.md, quickstart.md, and the contracts beginning with rest-api-v1.md

Tests: Required. The specification makes unchanged compatibility tests, contract tests, conformance tests, isolation tests, and UI tests release gates. Test tasks precede the implementation they verify.

Task format: [ID] [P?] [USn] Action in exact path — requirement(s); verifiable done condition

  • [P] means the task can run in parallel with adjacent [P] tasks because it changes different files and has no incomplete dependency.
  • [USn] maps directly to User Story n in spec.md.
  • A task is complete only when its stated test or observable condition passes. Do not mark implementation tasks complete from code review alone.

2026-08-18 Implementation Amendment: Independent AgentCore Slice​

This amendment supersedes any task below that edits ai_platform_engineering/dynamic_agents/ for the first implementation slice. Those tasks remain future migration backlog only. The current release gate is zero changed paths under that package.

  • T117 Create the independent Python package, locked dependencies, and non-root container in ai_platform_engineering/harness_engine/ — done when the service imports and its image has an isolated entrypoint.
  • T118 Add operator-owned AgentCore runtime aliases and reject arbitrary runtime identifiers in src/harness_engine/config.py, models.py, and adapters/agentcore.py — done when adapter tests prove the exact AWS request and bearer-token non-propagation.
  • T119 Add independent harness overlays, owned runs, ordered events, optimistic revisions, indexes, and TTL in src/harness_engine/repository.py — done for memory and Mongo repositories.
  • T120 Implement detached provider tasks and stable AgentCore runtimeSessionId ownership in src/harness_engine/coordinator.py — done when provider work is not scoped to an SSE request.
  • T121 Implement catalog, overlay, run, cancel, cursor replay, and SSE endpoints in src/harness_engine/main.py — done when internal credential and owner isolation tests pass.
  • T122 Add disconnect/replay E2E in tests/test_disconnect_replay_e2e.py — done when a run completes with no event subscriber and later replays from a cursor.
  • T123 Add stateless BFF proxies under ui/src/app/api/harness-engine/ with internal-service authentication and existing OpenFGA/conversation checks — done when Jest proves the user bearer is not forwarded and Last-Event-ID resumes a subscription.
  • T124 Add browser run/replay/subscribe helpers in ui/src/lib/harness-engine-session-client.ts — done when the public client contract separates start from subscription.
  • T125 Add harness-first Agent Creation UI selection and AgentCore runtime alias fields in DynamicAgentEditor.tsx, persisting only an independent overlay — done when existing editor tests and the AgentCore conditional-field test pass.
  • T126 Document service boundaries, configuration, and disconnect semantics in ai_platform_engineering/harness_engine/README.md and update this plan — done when limitations distinguish client disconnect survival from engine-process failover.
  • T127 Add deployment manifests and production secret wiring for the independent service without placing it in the default first-install profile — docker-compose.caipe-oss.yaml adds the service only to the explicit preview overlay, keeps Dynamic Agents untouched, builds the additive UI from the branch, and validates against the upstream Compose file.
  • T128 Add engine-process takeover/reconciliation for in-flight provider invocations — future reliability phase; browser/BFF disconnect survival is complete without it.
  • T129 Add stable provider thread identity plus validated W3C tracing storage/propagation for AgentCore runs — done when repeated conversation turns share a runtimeSessionId and the trace reaches the allowlisted runtime payload.
  • T130 Add the provider-neutral long-term MemoryBroker with explicit user/agent/organization scopes — future phase after the AgentCore session slice.

2026-08-18 Harness Gateway Vertical Slice​

  • T136 [US1] Persist an allowlisted execution_harness_id in the BFF-owned dynamic_agents record and default missing legacy records to Dynamic Agents — done when create/edit payloads retain the selected runtime without changing Dynamic Agents source.
  • T137 [US1] Implement deterministic target resolution in ui/src/lib/harness-gateway.ts — done when the selected marker routes to Harness Engine and missing/default markers preserve the existing DA proxy.
  • T138 [US3] Implement canonical-to-AG-UI and canonical-to-custom SSE translation in ui/src/lib/harness-gateway.ts — done when lifecycle and text frames are consumed unchanged by existing clients.
  • T139 [US1] Route the existing BFF start, invoke, cancel, and resume paths through Harness Gateway while preserving existing authentication, OpenFGA, and conversation checks — done without changes under ai_platform_engineering/dynamic_agents/.
  • T140 [US7] Add caller/agent/conversation-scoped active-run cancellation to Harness Engine without clearing the durable session binding — done when a follow-up run reuses the binding after cancellation.
  • T141 [US3] Add BFF compatibility coverage proving browser, Slack, and Webex AG-UI traffic routes through AgentCore and add Harness Engine cancel coverage — done when targeted Jest and pytest pass.
  • T142 Document the Harness Gateway boundary, routing marker, disconnect behavior, and current capability gates in the spec, plan, tasks, architecture, and Harness Engine README.
  • T143 [US3] Add provider-neutral interrupt resume to the adapter contract and Harness Gateway — blocked as a capability for the initial adapters, not silently delegated to Dynamic Agents.
  • T144 [US3] Add attachment normalization/storage references to Harness Engine — blocked as a capability for the initial adapters with explicit HARNESS_CAPABILITY_UNSUPPORTED responses.

Current gate: git diff main...HEAD --name-only | rg '^ai_platform_engineering/dynamic_agents/' produces no output; Harness Engine pytest/ruff, BFF Jest, editor Jest, ESLint, and TypeScript checks pass.

Workstream 1: Setup and Compatibility Freeze​

Purpose: Establish a measurable Dynamic Agents oracle before changing runtime behavior.

  • T001 [P] Capture the current generated OpenAPI document and route/status/header fixtures in ai_platform_engineering/dynamic_agents/tests/contract/fixtures/ and add the snapshot test in ai_platform_engineering/dynamic_agents/tests/contract/test_openapi_compatibility.py — FR-001, FR-006; done when the unmodified service matches the committed baseline.
  • T002 [P] Capture deterministic custom SSE and AG-UI golden frames for text, reasoning, tools, warnings, subagents, interrupts, resume, cancel, error, and finish in ai_platform_engineering/dynamic_agents/tests/contract/fixtures/streams/ with replay tests in ai_platform_engineering/dynamic_agents/tests/contract/test_stream_compatibility.py — FR-002; done when current output passes byte-for-byte.
  • T003 [P] Inventory current configuration fields, Mongo collections/indexes, cache keys, metrics, environment aliases, middleware outcomes, built-in tools, and error categories in docs/docs/specs/2026-08-17-harness-engine/compatibility-matrix.md — FR-001–FR-007; done when every row names a source path and baseline test ID.
  • T004 [P] Add deterministic fake-model/tool benchmark scenarios and a baseline runner in ai_platform_engineering/dynamic_agents/tests/performance/benchmark_compatibility.py — NFR-001, NFR-002, NFR-005; done when it records first-response, total-turn, event, memory, and cleanup baselines.
  • T005 Create contract/, conformance/, integration/, performance/, and security/ test packages under ai_platform_engineering/dynamic_agents/tests/ with shared neutral fixtures in ai_platform_engineering/dynamic_agents/tests/fixtures/ — FR-007, NFR-004; done when pytest collects each package without collection errors.
  • T006 [P] Record locked versions, licenses, transitive dependency diffs, network behavior, credential needs, and vulnerability review status for all four adapters in docs/docs/specs/2026-08-17-harness-engine/research.md — FR-011, FR-035; done when every adapter has an explicit enable/block decision.
  • T007 Amend .specify/ARCHITECTURE.md to state that all agents execute through Harness Engine and Deep Agents/LangGraph is the default compatibility harness — plan architecture gate; done when no living-architecture statement requires every agent to be LangGraph-native.

Checkpoint: The current implementation is frozen as a black-box oracle; no runtime behavior has changed.

Workstream 2: Foundational Contracts and Persistence​

Purpose: Add provider-neutral types, validation, persistence, and routes that block every execution story.

Critical: Runtime migration workstreams depend on these contracts. The independently implemented Phase 1 engine/gateway uses its own completed contract workstream and current-path release gates.

  • T008 [P] Add failing Pydantic compatibility tests for absent, valid, unknown, oversized, and unsafe harness/memory fields in ai_platform_engineering/dynamic_agents/tests/contract/test_harness_models.py — FR-005, FR-012–FR-015, FR-055–FR-057.
  • T009 Implement HarnessConfig, MemoryPolicy, descriptor, capability, validation-report, session-binding, lease, and canonical-interrupt models in ai_platform_engineering/dynamic_agents/src/dynamic_agents/models.py — FR-005, FR-009–FR-015, FR-044, FR-051–FR-057; done when T008 passes and legacy documents serialize unchanged.
  • T010 [P] Add canonical-event schema and lifecycle state-machine tests in ai_platform_engineering/dynamic_agents/tests/contract/test_canonical_events.py — FR-002, FR-016, FR-026, FR-029.
  • T011 Implement the canonical event union and validator in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/events.py — FR-026–FR-029; done when T010 rejects duplicate, late, malformed, oversized, out-of-order, and post-terminal events.
  • T012 [P] Add adapter/runtime protocol contract tests in ai_platform_engineering/dynamic_agents/tests/contract/test_harness_contract.py — FR-016–FR-025, FR-030.
  • T013 Implement immutable normalized inputs plus HarnessAdapter and HarnessRuntime protocols in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/contract.py — FR-016–FR-025, FR-030; done when no protocol exposes FastAPI, Mongo collections, bearer headers, or wire encoders.
  • T014 [P] Add registry, manifest-version, capability-level, certification, and partial-readiness tests in ai_platform_engineering/dynamic_agents/tests/contract/test_harness_registry.py — FR-008–FR-013, FR-035.
  • T015 Implement the static allowlisted registry and capability validator in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/registry.py and ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/capability.py — FR-008–FR-015; done when unsupported requirements fail before adapter construction and optional failures do not fail global readiness.
  • T016 [P] Add repository/index/idempotency/rollback tests for harness_sessions and harness_memories in ai_platform_engineering/dynamic_agents/tests/contract/test_harness_persistence.py — FR-004, FR-017, FR-052–FR-057.
  • T017 Implement Mongo collection handles, indexes, optimistic revisions, TTL behavior, and lazy creation in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/mongo.py and ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/session_store.py — FR-004, FR-017, FR-044, FR-052–FR-057; done when T016 and the rollback-readability cases pass without backfill.
  • T018 [P] Add catalog and validation REST contract tests in ai_platform_engineering/dynamic_agents/tests/contract/test_harness_routes.py — FR-009–FR-014, FR-069–FR-073.
  • T019 Implement GET /api/v1/harnesses, POST /api/v1/harnesses/validate, and sanitized per-harness health in ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/harnesses.py — FR-009–FR-014, FR-069–FR-073; done when reports include request ID, catalog revision, payload fingerprint, stable field path, step ID, constraints, and data-only fixes.
  • T020 Register additive harness routes and partial readiness in ai_platform_engineering/dynamic_agents/src/dynamic_agents/main.py and ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/health.py — FR-006, FR-035; done when existing route snapshots remain unchanged.
  • T021 [P] Add harness feature flags, default resolution, adapter capacity, and compatibility aliases to ai_platform_engineering/dynamic_agents/src/dynamic_agents/config.py and ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/config.yaml with tests in ai_platform_engineering/dynamic_agents/tests/test_harness_config.py — FR-005, FR-034–FR-037.
  • T022 [P] Add stable harness/adapter/execution-mode labels and redaction guards to ai_platform_engineering/dynamic_agents/src/dynamic_agents/metrics/agent_metrics.py with tests in ai_platform_engineering/dynamic_agents/tests/test_metrics_observability.py — FR-033, FR-058–FR-060, NFR-007.
  • T023 Run .specify/scripts/bash/check-prerequisites.sh --json --require-tasks --include-tasks, uv run ruff check, and the Workstream 1–2 pytest targets from ai_platform_engineering/dynamic_agents/ — all foundational requirements; done when the new contracts pass and the existing suite has no regression.

Checkpoint: Provider-neutral contracts and additive data/routes exist, but execution still uses the current runtime.

Workstream 3: User Story 1 — Deep Agents Compatibility (Delivery Phase 2)​

Goal: Execute existing agents through a deepagents compatibility adapter without changing clients, stored documents, streams, cache behavior, or rollback.

Independent Test: Run the entire unchanged Dynamic Agents suite and frozen black-box fixtures against the compatibility adapter with existing data and no harness backfill.

  • T024 [P] [US1] Add failing legacy-agent, cache, restart, and rollback tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_deepagents_compatibility.py — FR-004–FR-007.
  • T025 [P] [US1] Add LangGraph-chunk-to-canonical-event golden tests in ai_platform_engineering/dynamic_agents/tests/contract/test_deepagents_events.py — FR-002, FR-026–FR-029.
  • T026 [US1] Implement the compatibility adapter by delegating to the current AgentRuntime in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/deepagents_adapter.py — FR-005, FR-007, FR-016–FR-025; done when it does not fork current behavior.
  • T027 [US1] Implement LangGraph lifecycle translation in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/deepagents_events.py — FR-002, FR-026–FR-029; done when T025 matches all frozen frames.
  • T028 [US1] Add the common run/resume/cancel/restart lifecycle in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/runtime_coordinator.py — FR-001, FR-016–FR-018, FR-031–FR-032.
  • T029 [US1] Refactor ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/runtime_cache.py to cache HarnessRuntime objects while preserving keys, LRU, TTL, capacity, pending initialization, and shutdown behavior — FR-006, FR-034, NFR-005.
  • T030 [US1] Extract canonical-to-custom-SSE and canonical-to-AG-UI encoders under ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/stream_encoders/ — FR-002, FR-026–FR-029; done when T002 remains byte-compatible.
  • T031 [US1] Route start/invoke/resume/cancel/restart through the coordinator in ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/chat.py and provider-neutral interrupt operations in ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/conversations.py — FR-001–FR-003, FR-017–FR-018.
  • T032 [US1] Preserve absent-harness read behavior in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/mongo.py without writing a backfill — FR-004–FR-005; done when reads resolve deepagents and stored BSON remains unchanged.
  • T033 [US1] Add an explicit in_process compatibility/rollback switch in ai_platform_engineering/dynamic_agents/src/dynamic_agents/config.py and charts/ai-platform-engineering/charts/dynamic-agents/values.yaml — FR-006, FR-041; done when rollback changes execution mode without data migration.
  • T034 [US1] Run the unchanged Dynamic Agents pytest suite plus contract and conformance suites from ai_platform_engineering/dynamic_agents/ — SC-001, SC-002, SC-004; done with zero removed or weakened assertions.
  • T035 [US1] Run ai_platform_engineering/dynamic_agents/tests/performance/benchmark_compatibility.py against baseline and adapter — NFR-001, NFR-002, NFR-005; done when p95 and memory gates pass.

Checkpoint: This is the Deep Agents compatibility milestone in delivery Phase 2. It is not a prerequisite for the independent Phase 1 engine/gateway slice.

Workstream 4: User Story 4 — Add a Harness Without Rebuilding the Platform (P2)​

Goal: Make adapter certification repeatable without changing common routes, persistence, security, or encoders.

Independent Test: Register a neutral fake adapter, run the conformance kit, then remove it without editing shared request routing or client code.

  • T036 [P] [US4] Create neutral fake adapter/provider fixtures in ai_platform_engineering/dynamic_agents/tests/fixtures/fake_harness.py — FR-030–FR-032.
  • T037 [P] [US4] Add parametrized lifecycle conformance tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_runtime_lifecycle.py — FR-016–FR-018, FR-026–FR-032.
  • T038 [P] [US4] Add capability, manifest, schema, version-skew, health, and cleanup conformance tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_adapter_manifest.py — FR-009–FR-015, FR-030, FR-035.
  • T039 [P] [US4] Add tool, file, skill, subagent, interrupt, cancellation, and error conformance fixtures in ai_platform_engineering/dynamic_agents/tests/conformance/ — FR-018–FR-025, FR-031–FR-032.
  • T040 [US4] Add a registry-only fake adapter in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/testing_adapter.py gated to tests — FR-030; done when add/remove requires registry configuration only.
  • T041 [US4] Implement evidence-derived certification calculation in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/capability.py — FR-010–FR-011; done when adapters cannot self-declare certified.
  • T042 [US4] Add adapter-specific pytest markers and protected live-provider certification jobs in .github/workflows/ci-harness-conformance.yml — FR-011, NFR-004; done when ordinary CI uses fakes and certification CI records live evidence.
  • T043 [US4] Document the adapter authoring and certification workflow in ai_platform_engineering/dynamic_agents/README.md — FR-030–FR-032; done when a developer can add the fake adapter using only the published contract.

Checkpoint: A new adapter is bounded by contracts and evidence, not changes throughout the platform.

Workstream 5: User Story 3 — Same Chat Experience Across Harnesses (P1)​

Goal: Centralize tools, delegation, approvals, skills, files, and policy so every adapter produces the existing client experience.

Independent Test: Replay the portable baseline with the fake and Deep Agents adapters and compare normalized events, authorization decisions, side effects, and interrupts.

  • T044 [P] [US3] Add ToolBroker policy, credential isolation, retry, partial MCP, schema, approval, and result-bound tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_tool_broker.py — FR-019–FR-020, FR-032.
  • T045 [P] [US3] Add mixed-harness delegation, cycle, namespace, authorization, and nested-interrupt tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_delegation_broker.py — FR-022, FR-027.
  • T046 [P] [US3] Add portable attachment, skills, prompt context, middleware, and unsupported-content tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_portable_inputs.py — FR-021, FR-023–FR-025.
  • T047 [US3] Implement policy-preserving MCP and built-in execution in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/tool_broker.py by composing existing mcp_client.py, builtin_tools.py, and credential exchange — FR-019–FR-020, FR-032.
  • T048 [US3] Implement mixed-harness subagent dispatch in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/delegation_broker.py — FR-022, FR-027; done when each child receives independent authz, credentials, capacity, namespace, and trace context.
  • T049 [US3] Implement normalized prompt, skill, attachment, file, middleware, client-context, and user-context preparation in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/runtime_coordinator.py — FR-021, FR-023–FR-025.
  • T050 [US3] Implement provider-neutral pending interrupt persistence and approve/edit/reject/form resume mapping in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/session_store.py — FR-017–FR-018.
  • T051 [US3] Route Deep Agents tools and delegation through the common brokers in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/deepagents_adapter.py — FR-019–FR-022; done when existing credential and workflow delegation tests remain unchanged.
  • T052 [US3] Apply canonical size/order/sanitization bounds before wire encoding in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/events.py — FR-028–FR-029, NFR-007.
  • T053 [US3] Add client disconnect, cancellation-at-boundary, duplicate resume, and side-effect-race integration tests in ai_platform_engineering/dynamic_agents/tests/integration/test_run_races.py — FR-017–FR-018, FR-031–FR-032.
  • T054 [US3] Run all portable conformance cases against deepagents and the fake adapter — SC-003, SC-005; done when normalized lifecycle and policy outcomes match.
  • T055 [US3] Update ai_platform_engineering/dynamic_agents/SSE_EVENTS.md with canonical-to-existing-wire ownership while preserving the documented client contract — FR-002, FR-026–FR-029.

Checkpoint: Portable behavior is enforced by the engine and proven on two implementations.

Workstream 6: User Story 7 — Preserve Threads, Memory, and Traces (P1)​

Goal: Make pods disposable without losing thread continuity, governed cross-thread memory, or causal observability.

Independent Test: Persist thread and user memory, replace worker/control-plane replicas, resume exactly, and verify one sanitized trace across every boundary.

  • T056 [P] [US7] Add native-codec, durable-head, idempotency, interrupt-revision, restore-ownership, and uncertain-durability tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_thread_state.py — FR-051–FR-054, NFR-011.
  • T057 [P] [US7] Add user/agent/organization scope, authorization-before-search, provenance, approval, retention, conflict, injection, and deletion tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_memory.py — FR-055–FR-057, NFR-012.
  • T058 [P] [US7] Add W3C parentage, forged-context, baggage, redaction, provider-suppression, sampling, and collector-outage tests in ai_platform_engineering/dynamic_agents/tests/conformance/test_tracing.py — FR-058–FR-060, NFR-013–NFR-014.
  • T059 [US7] Implement binding-scoped native state codecs, compare-and-set durable heads, and generation checks in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/thread_state_store.py — FR-052–FR-054.
  • T060 [US7] Adapt existing LangGraph Mongo checkpoints to ThreadStateStore in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/deepagents_adapter.py — FR-004, FR-052–FR-054; done when existing checkpoints resume without rewrite.
  • T061 [US7] Implement commit ordering and explicit uncertain_durability handling in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/runtime_coordinator.py — FR-053, NFR-011; done when output-before-commit failures never auto-replay side effects.
  • T062 [US7] Implement authorized scoped memory CRUD/search, optimistic revisions, provenance, retention, and approval in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/memory_broker.py — FR-055–FR-057.
  • T063 [US7] Add narrow worker-facing state and memory facades in ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/internal_state.py and ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/internal_memory.py — FR-052–FR-057; done when binding/epoch/generation/authz are checked before storage access.
  • T064 [US7] Implement ingress trace-context sanitation, internal propagation, span topology, baggage allowlist, redaction, sampling, and bounded export in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/telemetry.py — FR-058–FR-060.
  • T065 [US7] Instrument authz, binding, coordinator, provider, state, memory, tools, subagents, validation, and encoding in ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/chat.py and common services — FR-058.
  • T066 [US7] Add randomized worker/control-plane crash-point recovery tests in ai_platform_engineering/dynamic_agents/tests/integration/test_state_recovery.py — SC-014; done after at least 1,000 deterministic failure points with zero acknowledged state loss.
  • T067 [US7] Add concurrent cross-scope memory isolation/conflict tests in ai_platform_engineering/dynamic_agents/tests/integration/test_memory_isolation.py — SC-015; done after at least 1,000 operations with zero unauthorized access or lost update.
  • T068 [US7] Add full-topology trace and leak-canary tests in ai_platform_engineering/dynamic_agents/tests/integration/test_trace_propagation.py — SC-016–SC-017.
  • T069 [US7] Implement the idempotent MongoDB index/collection migration and rollback verification described by mongodb-migration.md in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/mongo.py — FR-004, FR-052, FR-055.
  • T070 [US7] Run thread, memory, trace, migration, and existing Deep Agents persistence suites from ai_platform_engineering/dynamic_agents/tests/ — SC-014–SC-017; done when pod-local state is not required for any acknowledged durable outcome.

Checkpoint: External durable state, governed memory, and sanitized causal traces are ready for disposable workers.

Workstream 7: User Story 6 — Unique Sandbox Pods for Local Harness Sessions (P1)​

Goal: Run each local harness conversation epoch in its own claim-exclusive Agent Sandbox pod with no raw credentials and generation fencing.

Independent Test: Run 1,000 concurrent bindings and prove unique claim/Sandbox/pod UIDs, isolation, recovery, credential absence, egress policy, and stale-worker rejection.

  • T071 [P] [US6] Add worker protocol serialization, authentication, lifecycle, event, state, memory, tool, and cancellation tests in ai_platform_engineering/dynamic_agents/tests/contract/test_sandbox_worker_protocol.py — FR-043–FR-046.
  • T072 [P] [US6] Add claim lifecycle, readiness, reuse, hibernation, eviction, expiry, reconciliation, and fencing tests in ai_platform_engineering/dynamic_agents/tests/contract/test_sandbox_manager.py — FR-042–FR-045, FR-048–FR-049.
  • T073 [US6] Implement normalized run/resume/cancel and canonical streaming types in ai_platform_engineering/dynamic_agents/worker/protocol.py and the control-plane client in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/worker_client.py — FR-045.
  • T074 [US6] Implement the single-binding worker host, health, cancellation, cleanup, and adapter loader in ai_platform_engineering/dynamic_agents/worker/server.py — FR-043, FR-045, FR-048.
  • T075 [US6] Implement short-lived audience/binding/run/epoch/generation-scoped worker capabilities in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/worker_auth.py and worker verification in ai_platform_engineering/dynamic_agents/worker/auth.py — FR-044–FR-046.
  • T076 [US6] Implement allowlisted static profile resolution in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/sandbox_profiles.py — FR-042, FR-047; done when agent/request data cannot choose Kubernetes objects or images.
  • T077 [US6] Implement SandboxClaim creation, readiness, lease persistence, generation fencing, replacement, release, and orphan reconciliation in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/sandbox_manager.py — FR-042–FR-049.
  • T078 [P] [US6] Add pinned Agent Sandbox dependency/configuration and profile values to charts/ai-platform-engineering/Chart.yaml and charts/ai-platform-engineering/charts/dynamic-agents/values.yaml — FR-041–FR-049.
  • T079 [P] [US6] Add operator-owned SandboxTemplate/WarmPool manifests in charts/ai-platform-engineering/charts/dynamic-agents/templates/sandbox-profiles.yaml — FR-042, FR-047, FR-049.
  • T080 [P] [US6] Add worker RBAC, service, default-deny NetworkPolicy, explicit ingress/egress, security context, resource/PID/ephemeral limits, and no-token settings in charts/ai-platform-engineering/charts/dynamic-agents/templates/sandbox-security.yaml — FR-046–FR-047.
  • T081 [P] [US6] Add separately pinned Deep Agents, Claude SDK, and Strands worker images under ai_platform_engineering/dynamic_agents/worker/images/ — FR-041–FR-042; done when each image contains one adapter and uses an immutable non-root runtime.
  • T082 [US6] Route deepagents through the worker client and lease manager in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/deepagents_adapter.py, retaining explicit in-process fallback — FR-041–FR-045, FR-048.
  • T083 [US6] Add Kubernetes isolation, escape, egress, credential-canary, eviction, controller-outage, warm/cold readiness, and stale-generation tests in ai_platform_engineering/dynamic_agents/tests/integration/sandbox/ — FR-043–FR-050, NFR-008–NFR-010.
  • T084 [US6] Run ai_platform_engineering/dynamic_agents/tests/integration/sandbox/ across 1,000 bindings and render charts/ai-platform-engineering/charts/dynamic-agents/ — SC-010–SC-013; done when every local binding has distinct UIDs and old generations have zero accepted events, tools, or state writes.

Checkpoint: Deep Agents production execution is isolated per conversation epoch; the control plane contains no harness loop.

Workstream 8: User Story 2 — Select and Run Certified Harnesses Per Agent (P1)​

Goal: Add Claude Agent SDK, Strands Agents, and AgentCore behind the same contracts and certify each independently.

Independent Test: Create equivalent agents for every enabled harness, validate capabilities, and execute the common conformance suite through unchanged clients.

  • T085 [P] [US2] Add Claude SDK session-store, stream, tool, permissions/HITL, cancellation, and subprocess-cleanup conformance cases in ai_platform_engineering/dynamic_agents/tests/conformance/test_claude_sdk_adapter.py — FR-008, FR-016–FR-025.
  • T086 [P] [US2] Add Strands session-repository, hooks, stream, tool, interrupts, cancellation, and cleanup cases in ai_platform_engineering/dynamic_agents/tests/conformance/test_strands_adapter.py — FR-008, FR-016–FR-025.
  • T087 [P] [US2] Add AgentCore binding, deterministic remote-session, stream, inline-tool callback, IAM, timeout, cancel, clear, and collision cases in ai_platform_engineering/dynamic_agents/tests/conformance/test_agentcore_adapter.py — FR-008, FR-016–FR-025, FR-050.
  • T088 [US2] Implement the Mongo-backed Claude SDK session codec/store and worker adapter in ai_platform_engineering/dynamic_agents/worker/adapters/claude_sdk.py — FR-017–FR-025, FR-052–FR-054.
  • T089 [US2] Implement Claude SDK canonical translation, common ToolBroker facade, permission interrupts, scoped config directory, and cleanup in ai_platform_engineering/dynamic_agents/worker/adapters/claude_sdk.py — FR-018–FR-021, FR-026–FR-032.
  • T090 [US2] Implement the Mongo-backed Strands session repository and worker adapter in ai_platform_engineering/dynamic_agents/worker/adapters/strands.py — FR-017–FR-025, FR-052–FR-054.
  • T091 [US2] Implement Strands hooks/canonical translation, common tool facade, interrupts, cancellation, and cleanup in ai_platform_engineering/dynamic_agents/worker/adapters/strands.py — FR-018–FR-021, FR-026–FR-032.
  • T092 [US2] Implement the provider-managed AgentCore adapter, deterministic opaque session binding, stream translation, inline ToolBroker callbacks, IAM boundary, timeout, cancel, and clear in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/agentcore_adapter.py — FR-016–FR-032, FR-050.
  • T093 [US2] Register all adapters disabled/experimental with bounded JSON Schemas and contextual capability declarations in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/registry.py — FR-008–FR-015.
  • T094 [US2] Partition pending initialization and active capacity by harness in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/runtime_cache.py — FR-034, FR-036–FR-037.
  • T095 [US2] Add provider dependency/credential/model/region diagnostics with sanitized errors in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/harnesses/registry.py — FR-009, FR-012, FR-035.
  • T096 [US2] Run the full conformance suite per adapter and record evidence/status in docs/docs/specs/2026-08-17-harness-engine/compatibility-matrix.md — FR-011, SC-003; unresolved rows must leave the adapter experimental or blocked.
  • T097 [US2] Verify disabled/unhealthy adapter behavior and explicit no-silent-fallback policy in ai_platform_engineering/dynamic_agents/tests/integration/test_harness_availability.py — FR-035–FR-037, SC-005.

Checkpoint: Each non-default harness is independently available, experimental, blocked, or certified based on evidence.

Workstream 9: User Story 8 — Harness-Aware Agent Creation UI (P1)​

Goal: Preserve the existing editor steps while adapting models and fields to the selected harness without silent loss or unsafe browser authority.

Independent Test: Create, edit, and clone across every harness/catalog state; switch incompatible drafts; race catalog changes; and verify accessible blockers plus no partial writes.

  • T098 [P] [US8] Add strict HarnessConfig, HarnessDescriptor, FieldCapabilityReport, HarnessValidation, and AgentMemoryPolicy types in ui/src/types/dynamic-agent.ts with serialization tests in ui/src/types/__tests__/dynamic-agent-harness.test.ts — FR-061–FR-073.
  • T099 [P] [US8] Add BFF catalog/validation proxy tests in ui/src/app/api/dynamic-agents/harnesses/__tests__/route.test.ts and ui/src/app/api/dynamic-agents/harnesses/validate/__tests__/route.test.ts — FR-062, FR-064, FR-069–FR-073.
  • T100 [P] [US8] Add exact-payload, unsafe-field, stale-revision, no-Mongo-write, and no-OpenFGA-mutation tests in ui/src/app/api/dynamic-agents/__tests__/route-harness-validation.test.ts — FR-067, FR-069–FR-070.
  • T101 [US8] Implement sanitized catalog and draft-validation BFF proxies in ui/src/app/api/dynamic-agents/harnesses/route.ts and ui/src/app/api/dynamic-agents/harnesses/validate/route.ts — FR-062, FR-064, FR-069–FR-073.
  • T102 [US8] Filter/annotate model results by harness_id while preserving an incompatible stored model in ui/src/app/api/dynamic-agents/models/route.ts — FR-063.
  • T103 [US8] Extend the create/update mutable allowlist and strict validation for harness and memory in ui/src/app/api/dynamic-agents/route.ts and ui/src/app/api/dynamic-agents/agents/[id]/route.ts — FR-067–FR-070.
  • T104 [US8] Implement authorize-normalize-revalidate-fingerprint-write-OpenFGA ordering in ui/src/app/api/dynamic-agents/route.ts and ui/src/app/api/dynamic-agents/agents/[id]/route.ts — FR-069–FR-070; done when T100 proves all invalid/stale paths perform zero partial mutation.
  • T105 [P] [US8] Implement revisioned common draft state, per-harness transient option parking, and dirty snapshots in ui/src/hooks/use-agent-draft.ts with tests in ui/src/hooks/__tests__/use-agent-draft.test.ts — FR-065–FR-066, FR-071.
  • T106 [P] [US8] Implement catalog/model loading with AbortController and request-sequence guards in ui/src/hooks/use-harness-catalog.ts with tests in ui/src/hooks/__tests__/use-harness-catalog.test.ts — FR-062–FR-063, FR-071.
  • T107 [P] [US8] Implement debounced/immediate validation and fingerprint/report guards in ui/src/hooks/use-harness-validation.ts with tests in ui/src/hooks/__tests__/use-harness-validation.test.ts — FR-064, FR-069–FR-071.
  • T108 [P] [US8] Implement accessible harness radio cards and catalog states in ui/src/components/dynamic-agents/HarnessSelector.tsx with tests in ui/src/components/dynamic-agents/__tests__/HarnessSelector.test.tsx — FR-062, FR-073–FR-074, NFR-015–NFR-016.
  • T109 [P] [US8] Implement compatibility diff, explicit safe fixes, cancel/review behavior, and active-conversation confirmation in ui/src/components/dynamic-agents/HarnessCompatibilityDialog.tsx with tests in ui/src/components/dynamic-agents/__tests__/HarnessCompatibilityDialog.test.tsx — FR-065–FR-066, FR-072.
  • T110 [P] [US8] Implement native/emulated/unsupported/unavailable badges and unified step/field issue navigation in ui/src/components/dynamic-agents/HarnessCapabilityBadge.tsx and ui/src/components/dynamic-agents/CompatibilitySummary.tsx — FR-064, FR-074.
  • T111 [P] [US8] Implement portable memory controls in ui/src/components/dynamic-agents/AgentMemoryPolicyEditor.tsx — FR-068; done when only deployment-allowed scopes/policies/retention IDs can be selected.
  • T112 [P] [US8] Implement read-only persistence and tracing summaries in ui/src/components/dynamic-agents/ThreadPersistenceSummary.tsx and ui/src/components/dynamic-agents/ObservabilitySummary.tsx — FR-068; done when no storage, sandbox, secret, or protected-content capture control exists.
  • T113 [P] [US8] Implement first-party typed option panels and a static registry under ui/src/components/dynamic-agents/harness-options/ — FR-067; done when JSON Schema bounds values and no server-provided executable UI is rendered.
  • T114 [US8] Integrate harness-first selection and capability views into the existing five stable steps in ui/src/components/dynamic-agents/DynamicAgentEditor.tsx and preserve ui/src/components/dynamic-agents/deep-linking.ts IDs — FR-061–FR-068, FR-073–FR-074.
  • T115 [US8] Extend editor component tests under ui/src/components/dynamic-agents/__tests__/ for legacy default, filtered models, field states, blockers, unknown/unavailable harness, parked options, accessibility, and no silent reset — SC-018–SC-021.
  • T116 [US8] Add create/edit/clone, active-conversation, switch-losslessness, optional-outage, and stale-catalog Playwright coverage in ui/e2e/rbac/harness-agent-editor.spec.ts — SC-019–SC-021; done when browser and BFF race scenarios pass.

Checkpoint: Builders can safely select a harness without changing the familiar editor or trusting client-side compatibility decisions.

Workstream 10: User Story 5 — Operate, Compare, and Roll Back Harnesses (P2)​

Goal: Expose harness-aware health, telemetry, capacity, rollout, and containment without leaking sensitive data.

Independent Test: Run mixed traffic, inject one adapter/provider/controller/collector failure, and verify healthy harnesses continue while dashboards and errors identify the affected boundary.

  • T117 [P] [US5] Add harness/adapter/version/execution/result/latency/token/tool/interrupt/capacity metrics tests in ai_platform_engineering/dynamic_agents/tests/test_metrics_observability.py — FR-033–FR-037.
  • T118 [P] [US5] Add mixed-harness capacity, optional readiness, explicit fallback, timeout, and cleanup tests in ai_platform_engineering/dynamic_agents/tests/integration/test_harness_operations.py — FR-034–FR-037.
  • T119 [US5] Implement bounded low-cardinality metrics and health/capacity projections in ai_platform_engineering/dynamic_agents/src/dynamic_agents/metrics/agent_metrics.py and ai_platform_engineering/dynamic_agents/src/dynamic_agents/routes/health.py — FR-033–FR-036.
  • T120 [US5] Implement operator disable, per-harness admission, retryable capacity response, and fail-closed/explicit-fallback policy in ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/runtime_coordinator.py — FR-034–FR-037.
  • T121 [P] [US5] Add sanitized harness diagnostics to ui/src/app/api/dynamic-agents/health/route.ts and the existing admin runtime-health surface — FR-033–FR-036.
  • T122 [P] [US5] Add Harness Engine operational values and alerts to charts/ai-platform-engineering/charts/dynamic-agents/values.yaml and charts/ai-platform-engineering/charts/dynamic-agents/templates/ — FR-033–FR-037, NFR-008–NFR-014.
  • T123 [US5] Document health, capacity, traces, sandbox lifecycle, adapter disable, rollback, and uncertain-durability response in ai_platform_engineering/dynamic_agents/README.md — FR-033–FR-037, FR-048, FR-053.
  • T124 [US5] Run mixed-harness fault injection in ai_platform_engineering/dynamic_agents/tests/integration/test_harness_operations.py and verify one adapter failure causes zero failed runs or state leakage on healthy adapters — SC-005, SC-007.

Checkpoint: Operators can distinguish and contain every harness without inspecting sensitive content.

Workstream 11: Rollout, Hardening, and Release Gates​

Purpose: Prove the complete design, rehearse rollback, and update living documentation before production cutover.

  • T125 [P] Run uv run ruff check and the entire unchanged plus new pytest suite from ai_platform_engineering/dynamic_agents/ — FR-007, NFR-004; done with no weakened tests and at least 80% new-code coverage/100% contract-state and security-decision coverage.
  • T126 [P] Run nvm use, npm run lint, targeted Jest, Playwright harness-editor coverage, and npm run build from ui/ — FR-061–FR-074, SC-018–SC-021.
  • T127 [P] Render and validate charts/ai-platform-engineering/charts/dynamic-agents/ CRDs/RBAC/NetworkPolicy/profiles/pools and run docker-compose.yaml compatibility smoke tests — FR-006, FR-041–FR-050.
  • T128 Run ai_platform_engineering/dynamic_agents/tests/performance/benchmark_compatibility.py for compatibility latency, event translation, tracing overhead, warm/cold readiness, memory, and mixed-harness capacity — NFR-001–NFR-002, NFR-005, NFR-008, NFR-013.
  • T129 Run ai_platform_engineering/dynamic_agents/tests/security/ for authorization-before-provider, sandbox escape, credential canaries, cross-binding access, malformed events/context, SSRF, and telemetry leakage — NFR-003, NFR-007, NFR-009–NFR-010, SC-010–SC-017.
  • T130 Rehearse migration/index creation, rolling upgrade, in-process rollback, Dynamic Agents image rollback, and additive-data readability using docs/docs/specs/2026-08-17-harness-engine/mongodb-migration.md and quickstart.md — SC-002, SC-008.
  • T131 Implement and run shadow/canary controls in ai_platform_engineering/dynamic_agents/src/dynamic_agents/config.py and charts/ai-platform-engineering/charts/dynamic-agents/values.yaml, halting rollout on any compatibility/SLO gate — FR-036–FR-037, SC-001, SC-006.
  • T132 [P] Update evidence links and every adapter status in docs/docs/specs/2026-08-17-harness-engine/compatibility-matrix.md — FR-011; no planned row may be called certified.
  • T133 [P] Update ai_platform_engineering/dynamic_agents/ARCHITECTURE.md, ui/src/components/dynamic-agents/ARCHITECTURE.md, root AGENTS.md, and .specify/CHANGELOG.md to reflect the implemented boundaries — documentation constitution gate.
  • T134 Execute every scenario in docs/docs/specs/2026-08-17-harness-engine/quickstart.md and record commands/results in the release PR — all success criteria.
  • T135 Audit docs/docs/specs/2026-08-17-harness-engine/compatibility-matrix.md and confirm no service alias, public route, wire contract, existing collection, legacy field, or rollback artifact is removed; defer renaming/deletion to a separate spec — FR-001–FR-007, SC-004, SC-008.

Final checkpoint: Harness Engine can replace Dynamic Agents under the existing address, with Deep Agents compatibility proven, local sessions isolated, state/memory/traces durable and governed, optional adapters independently certified, UI creation safe, and rollback rehearsed.

Dependencies and Execution Order​

Delivery phases governs release order. Workstream numbers and task IDs below retain their historical identity and are not another phase sequence.

Technical dependencyDelivery meaning
Independent package/gateway/portable-contract/provider-lifecycle workstreamsPhase 1 foundation, with current-path review and release gates still required
Compatibility freeze and runtime/event/lifecycle workstreamsPrepare Phase 2 integration and preserve native behavior
UI catalog/validation foundationsPhase 1 capability truthfulness; Phase 2 reusable feature panels and switching
Native state, scoped credentials, worker protocol, sandbox lifecyclePhase 3 depends on proven state and identity boundaries; it does not block the existing experimental provider slice
Delegation contracts and child-run lifecyclePhase 4 builds on engine admission and authorized execution; native Deep Agents subagents remain a Phase 2 parity obligation
Operations, rollout and certificationApply to each selected delivery phase; certify only implemented capabilities

Stable backlog​

  • Existing task IDs and checkboxes are retained; this update does not revalidate historical completion claims.
  • Some original tasks target an earlier in-service layout. Reconcile their paths and contracts with the current engine package and Phase 2 sketches before implementation.
  • Validation tasks remain gates for the phase that implements the corresponding behavior. Do not infer that all target-architecture tests must pass to merge Phase 1.
  • Do not shadow live tool execution to compare backends; use recorded events, controlled fixtures or side-effect-free validation.

Implementation Strategy​

  1. Phase 1: complete the independent opt-in engine/gateway slice and its current-path correctness/review gates; preserve the Dynamic Agents default.
  2. Phase 2: integrate DeepAgentsAdapter and the pluggable builder through parity, identity, migration and rollback gates.
  3. Phase 3: implement external sandbox workers with SandboxManager retained inside Harness Engine.
  4. Phase 4: implement governed cross-harness delegation and child-run UI.

Service consolidation is a separate cutover after remaining APIs and operational responsibilities have replacements. Additional adapters and broker features qualify independently.

Completion Rules​

  • Write and observe each listed test fail before implementing its behavior.
  • Do not mark an adapter certified until every applicable compatibility-matrix row links to passing evidence.
  • Do not enable automatic fallback after a run may have produced output or side effects.
  • Do not run a production local harness in-process after its sandbox profile is certified.
  • Do not expose credentials, raw identities, arbitrary images/Kubernetes resources, or protected content to workers, agent configuration, UI, telemetry, or provider session IDs.
  • Do not acknowledge durable completion before native state and the binding head are committed.
  • Do not persist parked UI options, browser capability claims, or stale validation results.
  • Preserve the original Dynamic Agents rollback path until Workstream 11 completes.

Workstream 12: Independent Portable Abstraction Slice (Implemented)​

This workstream supersedes the earlier AgentCore-only overlay. It intentionally does not complete or modify tasks targeting ai_platform_engineering/dynamic_agents/.

  • T136 Define provider-neutral AgentBlueprint, HarnessDescriptor, capability levels, validation results, sessions, runs, and canonical events in ai_platform_engineering/harness_engine/src/harness_engine/models.py.

  • T137 Define broker boundaries for thread state, memory, tools, sandboxes, prompts, delegation, and telemetry in ai_platform_engineering/harness_engine/src/harness_engine/brokers.py.

  • T138 Implement the registry, catalog revision, normalized fingerprint, profile validation, and required-capability rejection in ai_platform_engineering/harness_engine/src/harness_engine/registry.py.

  • T139 Refactor AgentCore to the portable adapter contract with operator profiles, native session persistence, sanitized descriptor, and canonical events.

  • T140 Add a Claude Agent SDK adapter with operator profiles, safe option schema, native SDK session resume, text/usage translation, and injected-query tests.

  • T141 Replace overlay persistence with independent immutable agent versions, durable session bindings, runs, and replay events in memory and Mongo repositories.

  • T142 Refactor run coordination to pin sessions to agent versions, compile prompts, own provider tasks after disconnect, update native session IDs, and sanitize failures.

  • T143 Add catalog, draft validation, full agent save/get, detached run, replay, and cancellation APIs without importing or editing Dynamic Agents.

  • T144 Refactor the CAIPE editor and BFF to consume descriptors, render primitive JSON-Schema options, validate drafts server-side, and save independent blueprints.

  • T145 Add AgentCore, Claude SDK, immutable-version/session, ownership, disconnect/replay, schema-driven UI, and BFF tests.

  • T146 Document the portable contracts, session/thread/memory model, sandbox-pod target, UI contract, current limitations, and research in portable-abstractions.md and the component README.

  • T147 Move Harness Engine validation before the legacy Dynamic Agents write and provide atomic/recoverable dual-save coordination.

  • T148 Implement claim-exclusive Agent Sandbox pods, worker fencing, external checkpoint state, and tool/memory broker capability tokens for local SDK harnesses.

  • T149 Implement CAIPEAgentSessionManager for binding identity, ownership, immutable version pinning, durable lookup, provider event persistence, clear, and epoch rotation.

  • T150 Implement the provider-specific ProviderSessionManager contract and AgentCore deterministic plus Claude event-assigned strategies.

  • T151 Add authorized session-clear API/BFF routes and tests proving active-run cancellation, owner isolation, provider close, new binding identity, and adoption of the current agent version after clear.

  • T152 Add shared, accessible harness presentation metadata for LangChain Deep Agents, AgentCore, Claude SDK, and unknown adapters.

  • T153 Add a persistent chat identity header plus harness badges and deterministic agent colors in history and new-chat selection while preserving live/input/unread states.

  • T154 Add unit and component coverage for harness labels, deterministic themes, sidebar identity, and new-chat harness display.

  • T155 Resolve chat identity through a use-authorized metadata fallback and include the friendly agent name in the browser tab title.

Workstream 13: Per-agent AgentCore Harness Lifecycle​

  • T156 Extend sanitized AgentCore profiles with backward-compatible shared and operator-controlled per_agent modes.
  • T157 Add the private harness_provider_resources persistence contract for server-owned AgentCore identifiers and lifecycle state.
  • T158 Implement deterministic Harness naming/idempotency, CreateHarness/UpdateHarness readiness polling, reuse, failure sanitation, and DeleteHarness cleanup.
  • T159 Resolve per-agent Harness ARNs from the private repository during invocation; never accept or return them in portable blueprints.
  • T160 Add save conflict compensation and fail-closed provider deletion before BFF agent metadata removal.
  • T161 Add provisioning, reuse, private-identifier, invocation-resolution, API deletion, and BFF deletion-order tests.
  • T162 Configure least-privilege AgentCore control-plane, managed runtime endpoint/workload identity dependencies, and execution-role pass-through permissions in the preview deployment.
  • T163 Deploy to caipe-oss, create a neutral test agent, verify one new AWS Harness, execute multi-turn chat, delete the agent, and verify provider cleanup.