← Back to selected work

01 · Multi-agent delivery

Buzz, Hermes & agent-stack

A governed operating layer for coordinating Claude, Codex, Cursor, and specialist agents on real client work—with receipts, not vibes.

Evidence map

Durable receipts and explicit completion contracts

Source readbacks and route gates verify bounded work before completion.

  1. Capability route
  2. Isolated workspace
  3. Bounded worker
  4. Native verification
  5. Durable receipt
Boundary

People approve publication, destructive action, credentials, spend, and consequential judgment.

Constraint
Multiple AI runtimes can generate activity faster than a team can prove state, ownership, or what actually shipped.
Built
Buzz (Nostr team messaging) for channel coordination; Hermes as coordinator across bounded worker lanes; agent-stack as git-backed shared memory with reversibility tiers, completion contracts, and route-specific proof gates. Capability routing sends each task to the cheapest lane that can verify its own output.
Role
Designed and operates the stack on live ACFR and enterprise reporting engagements—scoped write packets, adopt-gates before production workbook changes, and durable session logs.
Human boundary
Humans retain approval for publication, destructive actions, credentials, spend, and consequential judgment. Automation proposes and routes; people decide.
Evidence
Thousands of production channel messages, worktree-isolated edits, native Workiva readbacks, and explicit [RECEIPT] / [GATE] completion contracts.
Scope
Practical implementation infrastructure for regulated delivery—not foundation-model or ML-research work.
Get in touch