HumanWire Agent Council

Inspiration

Important decisions rarely fail because teams cannot generate more text. They fail because coordination is messy: the right people are hard to reach, stakeholder concerns arrive late, evidence is confused with assertion, and a polished AI answer can look approved when no accountable person has approved it.

What it does

HumanWire Agent Council turns one executive agenda into durable decision work. Seven bounded Gemini specialists analyze market, finance, product, technical, risk, synthesis, and red-team concerns. Eight named AI stakeholders contribute role-specific evidence and authority constraints inside a clearly labeled Demo run.

The Council produces a review-ready decision brief that keeps confirmed facts, model inferences, red-team challenges, source records, an evidence digest, recommended action, and required authority separate. The final state remains Human Approval Required. Saved Firestore state survives refresh.

The public Demo run sends no external messages. Connected-organization transport is implemented but fails closed unless members are activated, routes are consented, and a private transport is ready.

How we built it

Gemini 3.5 Flash on Vertex AI performs bounded specialist analysis and synthesis.
Google ADK 2.7 orchestrates specialist roles and handoffs.
Cloud Run hosts one digest-pinned DecisionOS service.
Firestore stores tenant, mission, evidence, Council, decision, and audit state.
Firebase Hosting and Authentication provide the web experience and Google sign-in.
App Check is configured in monitored rollout; exact host, origin, CSRF, body, and schema controls protect mutation routes.
HumanWire—not the model—owns identity, evidence provenance, durable state, authorization, and final human approval.

The demo film also uses Veo 3.1 Fast for a clearly labeled visual guide and Lyria 3 Pro for its original instrumental score. Those assets explain the product; they do not substitute for real product proof.

Challenges

The hardest part was making autonomy visible without overstating it. We needed durable progress, safe retries, bounded model permissions, explicit human authority, truthful demo data, and a judge-readable record even when infrastructure or model calls fail.

Accomplishments

A real multi-agent decision workflow rather than a chat wrapper.
Seven visible Gemini specialists and eight named AI-stakeholder viewpoints.
An evidence-bound decision brief with sources, digest, red-team challenges, and required authority.
Durable refresh and replay from Firestore.
Google-native deployment with Cloud Run, Vertex AI, Firebase, and Firestore.
Test-driven privacy, authority, failure, and canonical-data boundaries.
A 1:52 public demo with real deployed product footage and visible Google Cloud proof.

What we learned

A useful enterprise agent needs more than better prompting. It needs durable state, authenticated identity, explicit authority, transparent handoffs, evidence provenance, and failure behavior that stays safe when models or infrastructure misbehave.

Built during the contest

HumanWire Agent Council is the Google-native competition entry built during the All Things Agentic submission period. It retains and discloses an earlier-in-period HumanWire coordination foundation. The Gemini/ADK Agent Council, Google runtime, Firebase sign-in, current cloud deployment, competition film, architecture, and submission materials are contest work.

Try it

Open the hosted workspace and sign in with Google.
Keep Demo run selected and start the prepared mission.
Wait for Decision brief ready.
Review Council, Decisions, and Evidence.
Refresh the mission URL to confirm the saved result and Human Approval Required state return.