crow-pet-agent-assembly

name	crow-pet-agent-assembly
description	Use when a task may need multiple agents, persistent role agents, cross-harness coordination, or generated subagents.

crow.pet Agent Assembly

Use this skill when Crow.PET needs to decide whether to create, select, or promote agents.

Crow.PET is the persistent super-agent. Skills are reusable procedures. Skill invocations are ephemeral. Agent instances are task or session workers. Agent profiles become persistent only when continuity matters or repeated use proves the role.

Workflow

1. Read the task record, user ask, linked SPEC.MD, work item, workflow definition, or stage handoff.
2. Decide whether a direct skill invocation is enough; avoid agents when one bounded skill can finish the task.
3. Choose agent lifetime: ephemeral, session, persistent, or personal.
4. Select or draft agent profiles only when role continuity, stable defaults, or repeated setup justify persistence.
5. Choose the collaboration pattern: Builder + Verifier, Scout + Specialist, Quorum, Local-first private pass, Cross-harness critique, or Parallel read, sequenced write.
6. Declare each agent's skills, context capsule needs, capability needs, policy limits, verification duties, and handoff contract.
7. Bind capabilities through the Capability Broker before selecting harnesses, models, tools, or factories.
8. Write a foundry brief when Pi or another factory should generate agents or subagents.
9. Route factory use through a factory shell agent that acts for the requesting LLM or harness inside Crow policy.
10. Keep factory output plan-only unless a live adapter path has Policy Gate Ask and an approval receipt.
11. Route agents through Crow-mediated records: Bus messages, subagent briefs, work items, stage handoffs, handoff reports, and execution traces.
12. Send useful events to the Ledger and learning candidates through the Loop.

Progressive Disclosure

Load workflows/assemble-agent-team.md when creating, selecting, or promoting agents.

Load ../../templates/foundry-brief.md only when a factory should generate agents or subagents.

Load ../../adapters/patterns/factory-shell-agent.md only when a factory should act on behalf of the current LLM or harness.

Load ../../templates/agent-profile.md only when selecting, drafting, or reviewing a persistent role agent.

Load ../../templates/stage-handoff.md when one agent's output becomes another agent's bounded input.

Load ../../templates/message-envelope.yaml and ../../adapters/patterns/bus.md when agents need typed communication.

Load ../../templates/execution-trace.yaml, ../../templates/context-delta.yaml, ../../templates/reflection-note.md, and ../../templates/curation-decision.md only when execution outcomes should feed the Loop.

Load ../../adapters/factories/pi.md only when Pi is the selected foundry.

Load ../../adapters/patterns/capability-broker.md and ../../adapters/patterns/tool-capability-profile.md only when agent roles declare semantic capability needs.

Load ../crow-pet-policy-gate/SKILL.md before external writes, private-context exposure, provider API billing, generated persistent agents, or tool permission changes.

Load contract.md and rubric.md only for acceptance or evaluation.

Dependency Edges

• delegates_to: crow-pet-spec-writer, crow-pet-work-queue, selected specialist skills, and generated task agents.
• uses_adapter: adapters/factories/pi.md, adapters/patterns/factory-shell-agent.md, Capability Broker, Tool Capability Profiles, Bus, Ledger, Loop, selected harness adapters, selected LLM adapters, and selected tool adapters.
• loads_context: task record, linked spec, work item, context capsule, candidate agent profile, stage handoff, and bounded Memory Garden notes.
• evaluates_with: contract.md, rubric.md, foundry brief, stage handoffs, handoff reports, and execution trace evidence.
• promotes_to: persistent agent profile, durable skill, playbook, rubric, policy, context delta, or Memory Garden note.

Rules

• Agents act; Crow remembers.
• Do not use free-form agent-to-agent chat as a durable primitive.
• Use the Bus for typed coordination and the Ledger for searchable evidence.
• Prefer ephemeral agents until continuity is justified.
• Promote persistent agent profiles through curated evidence, not convenience.
• Keep Pi as a foundry, not Crow's source of truth.
• Treat factory shell agents as bounded delegates, not OS shells, schedulers, memory stores, or policy authorities.
• Never silently fall back from local, private, or logged-in harness use to provider API billing.
• Keep generated agents disposable unless Self-Improvement promotes them.
• Require Policy Gate Ask and an approval receipt before live factory SDK, RPC, package install, provider API, or generated-subagent execution.
• Do not let a factory become Crow's scheduler, memory store, policy authority, or default runtime.