Bounded autonomy and operational KPIs

Harmonia uses bounded agency inside a durable workflow. Bedrock/Strands handles ambiguous media understanding and creative judgment. Deterministic code owns state transitions, validation, policy, cost reservation, idempotency, effects, and verification. A human owns consequential authorization. This hybrid design deliberately limits nondeterminism where correctness and accountability matter.

Authority by stage

Stage	Primary mode	What may vary	What cannot vary / stopping condition
collect_sources	deterministic workflow	provider availability and snapshot contents	sealed manifest, bounded source validation, and visible operator resolution for unsupported/unauthorized sources
extract_sources	agentic cognition inside deterministic workflow	Transcribe speech and Nova visual observations	normalized artifact schema, exact source digest, typed locators, budget, timeout, and visible failure contract
understand	agentic cognition	Nimi’s grounded moments and trend/meme angles	supplied evidence bounds, schema, and role policy
strategize	agentic cognition	Ryan’s four-week strategy and source-grounded briefs	strict context/evidence IDs, provenance validation, and a fail-closed allow-listed strategy-skill trace
awaiting_strategy_approval	deterministic authority	exact strategy decision	authenticated actor, canonical digest, expiry, and one-revision limit
plan	agentic cognition inside deterministic workflow	Temi’s skill-guided sequencing, cadence, supported channel/format, windows, deadlines, dependencies, and priorities using request-bound snapshot reads	deterministic snapshot assembly/persistence, strict trace and plan validation, canonical digests, deterministic eligible-item selection; no mutation or effect authority
draft	separate agentic Noni and Dara batch calls inside a deterministic bounded loop	typed multi-format content and structured editorial judgment	exact approved output plan; one original batch plus at most one issue-bound revision; exact lineage and evidence; every artifact must pass review; no effect authority
awaiting approval	human authority	approve or reject each proposed action	no external effect until a durable human decision exists
publish/export	deterministic workflow	provider response	official X post/thread or LinkedIn publication requires exact approval; internal artifact export is digest-bound and idempotent
verify	deterministic workflow	observed external/stored state	fresh official API read or dual-object byte reread; content packs require already-verified constituents
learn/memory	deterministic eligibility plus bounded context	eligible preference/outcome facts	evidence-linked workspace/brand scope; failure cannot erase an effect receipt

Chat and Telegram are control surfaces, not the autonomous core. The asynchronous pipeline continues in the background and every surface reaches the same policy and approval boundary. The dashboard’s operational projection is also a control surface, not workflow truth. It compiles DynamoDB jobs, approvals, connection health, resident attention, and unknown effects into a reconnectable needs-you inbox and compact job shells. Epoch-fenced pause, resume, and cancellation commands update durable control state; workers enforce that state when admitting new stage or effect claims. Resident autonomy extends that core with Heartbeat, Micro-reflection, Dream Cycle, and Wakeup Call. It does not extend model authority: verified live evidence and deterministic bounds decide what may be learned; protected changes require humans; reversible automatic tuning is restricted to an explicit allowlist. Replay and fixtures remain historical development inputs and cannot promote configuration. The autonomous core is the persisted job/effect-command pipeline driven by SQS. The full-deployment script configures an OIDC-authenticated EventBridge Scheduler invocation to run the lease-protected durable tick once per minute; stage-outbox recovery, scheduled work, proactive checks, and retention are isolated arms. Telegram interaction uses the authenticated webhook surface. The current web-only preview has not deployed this runtime.

Demo KPI contract

KPI	Definition	Evidence source	Current status
hands-off processing time	persisted job creation → first `awaiting_approval` event	DynamoDB job/events with one trace	pending authenticated run
approval latency	durable approval timestamp → effect receipt timestamp	decision + receipt	pending authenticated run
approved-output yield	approved actions with verified outcomes ÷ approved actions	decisions, receipts, verification	pending authenticated run
publish/verify success	verified real effects ÷ attempted approved effects	effect and independent readback/digest	pending authenticated run
duplicate-effect prevention	replay returns `already_applied` with the original receipt and creates no second effect	replay record + receipt/API readback	required by evidence verifier; pending authenticated run
cost per verified asset	reconciled observed model/media USD ÷ verified output count	immutable usage records + verification	pending authenticated run
operator steps eliminated	observed manual baseline steps minus actual Harmonia operator actions	time-stamped operator baseline worksheet + run events	pending a measured baseline; never inferred from code

The verifier computes consistency from redacted records; authenticated ECS Fargate, AgentCore Runtime, DynamoDB, SQS, model, effect, and verification exports establish provenance. Until that run exists, numeric KPI values remain explicitly pending rather than estimated or fabricated.

Authority and safety

Observability

Deployment

Authority by stage

Demo KPI contract

​Authority by stage

​Demo KPI contract

Authority by stage

Demo KPI contract