Agent skills, tools, prompts, and traces

Harmonia has a deterministic host coordinator and seven named operational specialists. Routing, context assembly, research and artifact variants, and resident Dream synthesis are separate registered tasks. The coordinator is host code; it is not a model that chooses its next stage. Skill guidance is method, never evidence. Search is a tool, not a skill. AgentCore Memory is advisory context, never authority.

Runtime inventory

Registration	Default model	Input → output	State key	Model-visible tools
`harmonia_intent_router`	Haiku 4.5	`IntentRoutingInput` → `IntentClassification`	`intent_classification`	None
`harmonia_context_assembler`	Haiku 4.5	`StrategyContextAssemblyInput` → `IntentStrategyContext`	`intent_strategy_context`	None
`nimi_analyst`	Nova 2 Lite	`AnalystInput` → `SemanticSourceAnalysis`	`source_analysis`	None
`nimi_research_analyst`	Nova 2 Lite	Same analysis contract	`source_analysis`	Exactly one of `nimi_gateway_search` or `nimi_agent_search_agent`, selected before inference
`ryan_strategist`	Sonnet 4.6	`StrategistInput` → `StrategistResult`	`strategist_result`	None
`ryan_research_strategist`	Sonnet 4.6	Same strategy contract	`strategist_result`	`ryan_gateway_search`
`temi_editorial_planner`	Sonnet 4.6	`EditorialPlannerInput` → `EditorialPlan`	`editorial_plan`	None; the host supplies the planning snapshot
`noni_copywriter`	Sonnet 4.6	`CopywriterInput` → `ContentDraft`	`copywriter_draft`	`search_verified_publications`
`noni_artifact_producer`	Sonnet 4.6	`ArtifactProductionInput` → `SemanticArtifactWireDraft`	`semantic_artifact_draft`	None
`dara_editor`	Sonnet 4.6	`EditorialReviewInput` → `EditorialAssessment`	`editorial_assessment`	None
`dara_artifact_editor`	Sonnet 4.6	`ArtifactReviewInput` → `SemanticArtifactReview`	`semantic_artifact_review`	None
`maya_presenter`	Haiku 4.5	`UiContext` → `SurfacePlan`	`surface_plan`	None
`nova_liaison`	Haiku 4.5	Host-validated `LiaisonInput` → `LiaisonAnswer`	`liaison_answer`	Six scoped reads listed below
`nova_context_answer`	Haiku 4.5	`LiaisonInput` → `ContextRecordAnswer`	`context_record_answer`	None; answers only from the supplied context record
`harmonia_dream_synthesizer`	Sonnet 4.6	`DreamCycleInput` → `DreamCycleOutput`	`dream_cycle_output`	None

Exact model IDs are listed in Bedrock models. SourceAnalysis is the host-bound persisted analysis, not the raw semantic model result.

Preloaded method guidance

Role	Method module	Preloaded skill
Nimi	`nimi_skills.py`	`nimi-analysis-skills`
Ryan	`ryan_skills.py`	`ryan-strategy-skills`
Temi	`temi_skills.py`	`temi-editorial-planning-skills`
Noni	`noni_skills.py`	`noni-writing-skills`
Dara	`dara_skills.py`	`dara-editing-skills`
Nova	`skills_runtime.py`	`trend-scan`, `signal-watch`, `engagement-insights`, `posting-schedule`, `job-status`

The host compiles method text into each invocation and bootstraps activation provenance. The runtime does not register load_skill or load_skill_resource. Shared handoff/recovery instructions accompany every specialist. Method files and their hashes are not factual source evidence. Nova exposes fetch_trend_signals, search_trend_signals, get_engagement_insights, get_operator_feed, get_job_status and suggest_posting_windows. The host validates returned evidence and permits only a bounded retry of the same transient read.

Request and tool enforcement

build_agent_team() creates SpecialistDefinition values. HarmoniaCoordinator.select() chooses the exact allowed definition from host state. LocalStrandsTeamRuntime constructs a fresh strands.Agent using BedrockModel, structured_output_model, AuthorityHooks, callback_handler=None and retry_strategy=None, then calls invoke_async(). BeforeModelCallEvent records hashes of the actual request and enforces the paid-operation gate. BeforeToolCallEvent validates tool membership and the exact research request, consuming one-shot research authority before dispatch. AfterToolCallEvent records typed evidence. The native structured-output tool is a cognition contract, not permission for an external action. Gateway Web Search and Bedrock Knowledge Base retrieval use native AWS transports. Provider evidence is normalized at the host boundary. No Google built-in grounding tool is registered, and Noni has no public-web search registration.

Persistence and evidence status

The runtime returns state keys; worker code validates and persists accepted results. DynamoDB operations, approval digests, effect claims and independent readback own execution truth. A model cannot authorize an effect or create an execution receipt. The registrations and request contracts have local tests. Authenticated AgentCore, Bedrock, Gateway and Memory execution in this AWS edition remains unverified. See agent/harmonia_agent/agents.py, team_runtime.py, agentcore_app.py and runtime_instructions.py.

Research and strategy

Planning and production

Experience and liaison

Agent runtime

Runtime inventory

Preloaded method guidance

Request and tool enforcement

Persistence and evidence status

​Runtime inventory

​Preloaded method guidance

​Request and tool enforcement

​Persistence and evidence status

Runtime inventory

Preloaded method guidance

Request and tool enforcement

Persistence and evidence status