Skip to main content

Agent teams

Use the agent-team handbook to discover or create a team, select its project manifest, assign role ownership and run the revisioned task lifecycle. It is the canonical guide served directly from docs/guide; this entry links to it rather than copying its procedures.

Local coordination requires Node.js 22+ and the complete creator payload. Native Codex, Claude Code, OpenCode, Kimi and DeepSeek execution remains subject to each installed harness's discovery and permissions. Export and project installation prepare definitions; they do not launch workers or activate UAR/BossFang teams. Model and memory services are separate optional capabilities.

Start at your level​

You need to…Read
Turn an outcome into a small teamUser and lead responsibilities
Reuse an existing project teamCreate, select or adopt
Give a role work it can safely executeOwnership and dispatch example
Track progress or recover interrupted workRevisioned lifecycle and canonical recovery
Move work to another role, harness or projectAccepted handoffs and project requests
Set a model, effort or fallbackModel selection, native limits and fallback decisions
Share lessons with the right audiencePrivate, team, project and broader scopes
Decide whether services or Companion are neededServices for team work

Procedures and request examples​

The full pack includes broader learning integrations; mini retains its Node-only portable runtime with surreal-memory and liter-llm as its two optional services. Local team state works without them. Source examples and schema receipts do not establish live native acceptance, successful inference or deployed execution.

Communicate and transfer context​

Follow the canonical guide for same-team updates, context transfer and acceptance, cross-project intake and scoped learning. Use the two-role executable request sequence for the precise creation, revisions, accepted review transfer and completion fields.

Full has explicit card publication/discovery and issue request/intake commands; mini currently requires authorized manual cross-project delivery. An imported request is pending triage, not accepted work. Issue writes need explicit authorization and gh; local handoff needs no service or remote chat API. Full's learning writer and optional Cortex mirror are separate from mini's Claude-only file-tier delivery. Neither a scope label nor a mirror admission receipt proves private storage, delivery or recall. Inspect actual packaged and installed capabilities before using these source examples.

Select and record an actual model route​

Use the canonical model guide for declared-versus-observed evidence, policy precedence, native reasoning limits, budgets, explicit fallback and reviewer independence. The full model request sequence shows discovery, effective selection and persisted native intent. Record each canonical assignment through task model assignments.

KBD chooses task fit first; the helper filters constraints then sorts eligible candidates by known prices. Neither selection nor export switches a running worker. liter-llm inference needs a separate documented worker for tool execution. Unknown metadata can leave the route unresolved; no automatic fallback or universal per-role reasoning control is promised. A same-family critic does not satisfy different-family KBD review, even if its alias or effort differs.