Give it a mission. Get back the work — and the proof.

chief_supervisor → crew_manifest

The Chief Supervisor queries the model registry and tool gateway to compose a crew:
  • Each team gets a named Team Supervisor
  • Each Team Supervisor is assigned a set of Expert Agents
  • Models are selected from the registry by capability tag
  • Tools are granted by the tool gateway per agent role

team_supervisor → task_assignment

The Team Supervisor decomposes the team objective into tasks and routes each to the appropriate Expert Agent:
  • Tasks are typed (research | draft | review | benchmark | …)
  • Each task carries a sub-budget inherited from the mission budget
  • Routing decisions are logged to the ledger

expert_agent → artifact: WorkProduct

Expert Agents publish typed work artifacts to the shared blackboard:
  artifact {
    type:      "design_doc" | "benchmark_result" | "review_note" | …
    producer:  agent_id
    content:   string | object
    hash:      sha256(content)
    timestamp: iso8601
  }

Downstream agents consume artifacts by type, not by producer name.

expert_agent → tool_gateway → external_tool

All tool calls are routed through the tool gateway:
  request {
    tool:   "web_search" | "code_exec" | "file_read" | …
    args:   object
    budget: { calls_remaining: int }
  }

The gateway enforces per-tool rate limits, logs every call, and blocks tools not in the agent's grant list.

team_supervisor → human_checkpoint (if autonomy_level ≠ autonomous)

When an agent reaches a decision point outside its autonomy level, the Team Supervisor raises a clarification request:
  • The mission pauses at the checkpoint
  • A human operator reviews and responds
  • The response is logged and the mission resumes
  • Timeout triggers escalation to the Chief Supervisor

expert_agent → team_supervisor → next_agent

Typed handoffs carry the artifact reference and a structured context object:
  handoff {
    from:     agent_id
    to:       agent_id | "team_supervisor"
    artifact: artifact_id
    context:  { summary: string, open_questions: string[] }
    hash:     sha256(artifact + context)
  }

The hash chain links every handoff to the previous event in the ledger.

expert_agent → artifact: FinalDraft

The final draft artifact is published to the blackboard and flagged for QA review:
  artifact {
    type:      "final_draft"
    producer:  agent_id
    content:   string | object
    hash:      sha256(content)
    ready_for_qa: true
  }

qa_reviewer → verdict: pass | fail | escalate

The QA Reviewer runs on a different model family than the builder:
  • Receives the final draft and the definition of done
  • Has no access to the builder's reasoning or intermediate artifacts
  • Returns a structured verdict:
      { result: "pass" | "fail" | "escalate", findings: string[] }
  • A fail triggers a revision cycle (up to max_revisions)
  • An escalate raises a human checkpoint

chief_supervisor → evidence_pack

On QA pass, the Chief Supervisor seals the mission:
  evidence_pack {
    mission_id:   uuid
    plan:         MissionPlan
    artifacts:    Artifact[]
    tool_calls:   ToolCall[]
    handoffs:     Handoff[]
    qa_verdicts:  QAVerdict[]
    ledger_root:  sha256(chain_tip)
    sealed_at:    iso8601
  }

The evidence pack is signed and stored. The ledger root allows offline verification of the entire run.

When the Chief Supervisor cannot resolve an ambiguity from the proposal alone, it raises a clarification interview — a structured dialogue with the operator that resolves the ambiguity before the mission begins.

The interview is bounded: a maximum of five rounds, each with a specific question and a set of acceptable answer types. If the ambiguity is not resolved in five rounds, the mission is declined with a reason.

clarification_interview {
  round:    int (1–5)
  question: string
  options:  string[] | null  // null = free text
  context:  { proposal_excerpt: string, ambiguity: string }
}

Clarification responses are logged to the ledger as part of the mission record. The final accepted proposal — including all clarification responses — is included in the evidence pack.