ADK agent platform

Harmonia uses Google ADK for bounded judgment inside a deterministic asynchronous workflow. Agent Engine hosts cognition; it does not own durable workflow state or external-effect authority.

Harmonia

Owns intent routing, authoritative reads, handoffs, bounded correction, and escalation.

Nimi & Ryan

Produce grounded analysis, strategy, and content plans.

Temi

Agentically operationalizes an approved Ryan strategy; deterministic code persists the plan and selects one item.

Writing & review

Deterministic code runs separate typed Noni and Dara calls for the selected item.

Team topology

Role	Responsibility	Allowed context	Prohibited authority
Harmonia	intent routing, exact specialist selection, host-owned handoff construction, authority-read orchestration, bounded correction, and escalation	typed invocation state plus authenticated host reads	invent identifiers, digests, tool results, approval, receipts, verification, or provider outcomes
Nimi	source evidence analysis: exact moments and qualified source/context/performance/memory angles	typed transcript/frame evidence, verified performance, eligible Memory Bank facts, one bounded method skill, and optional request-bound grounded context	use skill prose as evidence, invent evidence or trends, define strategy, write final copy, approve, schedule, publish, or treat memory/search as authority
Ryan	agentic strategy and source-grounded content briefs	typed Nimi analysis, company/campaign context, verified performance, eligible Memory Bank facts	write final copy, mutate calendars, approve, publish
Temi	agentic editorial operationalization: sequencing, cadence, supported channels/formats, windows, deadlines, dependencies, and priorities	exact human-approved Ryan strategy, Nimi evidence, one immutable Firestore planning snapshot, local method skill, and request-bound read views	invent strategy/evidence, search, read outside the snapshot, write final copy, approve, mutate external calendars, schedule externally, publish
Noni	platform-native writing and revision	one selected Temi item, its exact Ryan brief, referenced Nimi evidence, local writing guidance, verified prior publications, and brief-scoped public sources	redefine strategy, search trend feeds, access Memory Bank, inspect the rest of the plan, create facts, approve, publish
Maya	reference-only interface composition	exact typed job/entity catalog; deterministic host hydration from Firestore	invent IDs or state, emit lifecycle placeholders, expose payloads, approve, publish, verify, or mutate workflow
Dara	complete seven-dimension assessment and bounded correction issues	exact Noni draft, immutable production input, and allow-listed editing-method guidance	cite skill prose as evidence, author workflow metadata, rewrite copy, replace evidence, request a third pass, or create authority
Maya	A2UI layout selection	typed entity references and UI context	emit executable UI or mutate records
Nova	read-only operational answers	allow-listed tools and skills	write Firestore, retry jobs, approve, publish

Managed invocation lifecycle

Reconstruct a typed snapshot

The worker reads authoritative Firestore state and creates a serializable invocation seed.

Create or resume a namespaced session

Agent Engine receives a deterministic session identity derived from the user and operation. Existing managed state is read without mutating the caller’s payload; a missing session is created once and reused after process restarts.

Stream typed deltas

ADK output_key values and event state_delta records carry specialist handoffs. Required keys are validated before use.

Persist validated results

Deterministic routes write accepted results to Firestore. Session state alone cannot advance the workflow.

Reconcile interrupted streams

If the managed event stream terminates after state was durably committed, Harmonia rereads the session and accepts only validated state that differs from the invocation seed. Missing or partial output remains a visible failure.

Agent state, chat history, and Memory Bank are context—not authorization. Only durable policy, approval, claim, receipt, and verification records control effects.

Invalid specialist output crosses the coordinator boundary as a structured AgentContractError with a role, stable code, safe operator message, and optional schema path. The failure pipeline retains these fields in its typed envelope and durable activity event; model output and validation payloads never become operator-facing error text.

Coordination and bounded course correction

Harmonia is the deterministic ADK root. The host places the exact requested specialist in session state; the root resolves that registered child directly instead of asking a model to invent or rewrite an agent transfer. Every specialist receives a host-created harmonia.handoff/v1 envelope and acknowledgement bound to the specialist, operation ID, input digest, and output contract. Specialists return semantic judgment only. Application code computes canonical identifiers and digests, materializes artifacts, records actual callback/tool evidence, and decides whether an output can advance. A missing authority read, invalid evidence reference, malformed structured result, or incomplete acknowledgement produces a safe repair code. Harmonia can run at most two issue-specific corrections against the original trusted input. A third invalid result becomes agent_output_repair_exhausted and is escalated; the coordinator does not broaden authority, change the goal, or fabricate a successful handoff. Structured output remains enabled for tool-capable Gemini specialists. Tool calls and output schemas are complementary boundaries: callbacks record real tool execution, while the schema constrains the final semantic result. Skill activation records are identified separately and are never presented as tool results.

Activity and telemetry

Every specialist invocation emits a strict metadata-only log, trace, and metric projection with tenant, job, operation, agent, stage, model, timing, usage counts, outcome, and real W3C trace/span IDs. Tool activity is projected only when ADK supplies real trace context. Unknown fields fail validation, and the schema cannot carry prompts, model responses, transcripts, drafts, credentials, or provider bodies. The projection powers Monitoring → Agent activity. Native ADK OpenTelemetry signals continue to Cloud Logging, Cloud Monitoring, and Cloud Trace; the projection is a bounded operator view, not a replacement for those backends and never an authority source.

Versioned ADK skills

Nimi loads exactly one nimi-analysis-skills skill and at least one relevant allow-listed analysis reference. This filesystem skill contains method guidance only. Dara loads exactly one dara-editing-skills skill and at least one approved editing reference before review. The seven references route editorial triage, grounding and claims, structure and clarity, brief/voice/audience fit, platform/CTA usability, safety and inclusion, and bounded feedback. Dara has no search, Memory Bank, workflow, approval, or effect tool. Runtime callbacks record the actual loader sequence and deterministic validation rejects missing, duplicate, out-of-order, or unapproved resources. The guidance can shape editorial judgment but cannot serve as factual evidence, a brand constraint, or authority. Noni loads exactly one noni-writing-skills skill and at least one approved writing reference before drafting. Temi loads exactly one temi-editorial-planning-skills skill and at least one of six allow-listed planning references. Its separate read-only tools expose commitments, capacity, asset readiness, verified posting-window observations, calendar projection state, and blocked dependencies from the exact immutable planning snapshot. The trace must load references before data reads and every read must match the request’s snapshot ID. Skill prose is methodology and never evidence.

Runtime research tools

Search and internal-data reads are tools, not skills. Nimi’s optional native Google Search and Vertex AI Search tools run in isolated request-bound child agents. The runtime validates a separate research trace and native grounding metadata before persistence. Noni has two separate read capabilities: search_verified_publications for tenant-scoped posts with applied receipts and successful verification, and native ADK google_search through an isolated brief-bound research agent. It has no trend-feed, Memory Bank, approval, or effect tool. The eleven references cover thought leadership, hooks and introductions, MECE structure, case studies, storytelling, bad-content diagnosis, persuasion, outlining, titles and headlines, convincing content, and narrated short-form video scripts. Skill guidance shapes writing but never counts as evidence for a factual claim. Research-derived claims and URLs must cite the live tool evidence ID recorded in the actual trace. Nova loads exactly one skill before any data tool. The runtime records the real skill/tool envelopes in managed ADK state and rejects answers whose selected skill, tool sequence, retry behavior, status, errors, claims, or evidence IDs do not match that trace. Skill prose guides reasoning; deterministic validation enforces the boundary.

Skill	Purpose	Allowed tools
`trend-scan`	find grounded public topics	trend signal fetch/search
`signal-watch`	explain current signals and operator feed	trend signals and operator feed
`engagement-insights`	summarize measured post performance	engagement insights
`posting-schedule`	suggest evidence-based UTC windows	engagement insights, operator feed, posting windows
`job-status`	answer scoped operational status questions	job status

See Agent tool contracts for the exact envelope and permission model.

Research and strategy

Planning and production

Experience and liaison

Agent runtime

ADK agent platform

Harmonia

Nimi & Ryan

Temi

Writing & review

Team topology

Managed invocation lifecycle

Coordination and bounded course correction

Activity and telemetry

Versioned ADK skills

Runtime research tools

Harmonia

Nimi & Ryan

Temi

Writing & review

​Team topology

​Managed invocation lifecycle

​Coordination and bounded course correction

​Activity and telemetry

​Versioned ADK skills

​Runtime research tools

Team topology

Managed invocation lifecycle

Coordination and bounded course correction

Activity and telemetry

Versioned ADK skills

Runtime research tools