@ai2070/crew

Deterministic, replayable, bus-agnostic crew orchestration for AI agents.

A crew is a graph of typed roles and agents that takes an input through a sequence of phases and produces an output. This library does the orchestration — phase ordering, parallelism, voting, fault recovery, nested crews, scoped memory — without making any decisions about how agents actually run. Plug it onto whatever transport you want.

What it is

A state machine, not an executor. The library:

• Validates a crew shape and agent counts (Zod v4) and materializes a virtual graph.
• Emits lifecycle events (agent.step.requested, vote.resolved, fixer.invoked, nested.crew.completed, …) that callers publish onto a bus.
• Consumes responses (agent.step.completed, agent.step.failed, agent.stream.chunk) and advances accordingly.
• Tracks parallel agents, dedups duplicate deliveries, manages timeouts, routes faults to fixer agents, and runs nested crews with optional memory isolation.

It is not:

• A model runner. The library never calls L0 / OpenAI / Anthropic / anything.
• A bus. There's no preferred transport (Net, EventEmitter, Kafka, in-process queue, anything works).
• A persistence layer. Snapshots are values; storage is the caller's job.
• An async framework. Hooks are sync; the loop is sync; everything is (events in) → (events out).

Architecture

                ┌────────────────────────┐
                │     crew library       │
                │  (state machine only)  │
                └────────────────────────┘
                      ▲              │
             inbound  │              │ outbound
             events   │              │ lifecycle events
                      │              ▼
                ┌────────────────────────┐
                │         bus            │  ← Net, EventEmitter, …
                │  (caller's choice)     │
                └────────────────────────┘
                      ▲              │
                      │              ▼
                ┌────────────────────────┐    ┌──────────────┐
                │     agent worker       │ →  │      L0      │
                │   (caller's code)      │    │   (model)    │
                └────────────────────────┘    └──────────────┘

The library is the box at the top. Everything else is the caller's wiring.

Install

npm install @ai2070/crew zod

Optional peers:

npm install @ai2070/memex     # for scoped per-agent memory

Quick Start

import {
  CrewShapeSchema,
  CrewAgentsSchema,
  buildCrewGraph,
  createCrewSession,
  systemClock,
} from "@ai2070/crew";

// 1. A crew is two JSON blobs: the shape (roles, permissions, capabilities)
//    and the counts (how many of each role).
const shape = CrewShapeSchema.parse({
  schema_version: "1.0",
  name: "RESEARCH_CREW",
  roles: [
    {
      role: "caller",
      capabilities: { thinking_allowed: false },
      permissions: { talk_to: ["fixer"], delegate_to: [], escalate_to: [], invite: [] },
      first_input: true,
      final_output: true,
    },
    {
      role: "merc",
      capabilities: { thinking_allowed: true, model: "claude-sonnet-4-6" },
      system_prompt: "You are a research operative. Gather facts efficiently.",
      permissions: { talk_to: ["caller"], delegate_to: [], escalate_to: ["fixer"], invite: [] },
      amount: 4,
    },
    {
      role: "fixer",
      max_allowed: 1,
      capabilities: { thinking_allowed: true },
      activation: { on_fault: true, on_stall: true },
      permissions: {
        talk_to: ["merc", "caller"],
        delegate_to: ["merc"],
        escalate_to: ["caller"],
        invite: [],
      },
    },
  ],
});

const counts = CrewAgentsSchema.parse({
  schema_version: "1.0",
  name: "RESEARCH_CREW",
  agents: [
    { role: "caller", amount: 1 },
    { role: "merc", amount: 4 },
    { role: "fixer", amount: 1 },
  ],
});

// 2. Build the graph (validated, lint-checked, deterministic).
const graph = buildCrewGraph(shape, counts);

// 3. Create a session.
const session = createCrewSession({
  crewId: "research-1",
  graph,
  clock: systemClock(),
});

// 4. Drive the session against your bus. The optional `task` argument is
//    session-level metadata — emitted once on `crew.started`, captured in
//    snapshots, NOT propagated to per-step requests.
const initial = session.start("What sleep patterns do birds have?", {
  description: "Research bird sleep patterns",
});
for (const e of initial) bus.publish(e);

bus.subscribe("agent.step.completed", (e) => {
  for (const out of session.deliver(e)) bus.publish(out);
});

// Periodically tick to fire timeouts.
setInterval(() => {
  for (const out of session.tick(Date.now())) bus.publish(out);
}, 1000);

The bus consumer (somewhere else in your stack) listens for agent.step.requested, runs whatever model/prompt logic it wants (e.g. calls L0), and publishes agent.step.completed back. The crew session never knows.

Mental Model

• The shape declares the topology — which roles exist, how they relate, who can talk/delegate/escalate to whom, what activates fault recovery.
• The session emits requests — when a phase starts, the session emits agent.step.requested for every agent in that role with a self-contained role snapshot, the input, and the optional memex context. Workers handle them however.
• The session waits for responses — every request has a deterministic correlationId. Responses match by id. Phases wait for all agents to resolve before voting.
• Voting picks one output per phase — first_valid (default), majority, unanimous, or weighted_consensus. The resolved value becomes the input to the next phase.
• The cascade is one-way — phase N+1 only sees phase N's resolved output as its input. The original rootInput only reaches the caller (entry-point role). Each role brings its own memex view and own system_prompt. Workers compose the model prompt from those three pieces.
• Faults route to fixer — if any role has activation.on_fault, an agent's failure spawns a fixer step whose response substitutes for the failed agent's slot in voting.
• Nested crews are recursive — a role can declare nested_crew, which causes its agents to spawn inner sessions instead of going to the bus. Memory can be soft-isolated (shared adapter) or hard-isolated (forked, merged on completion).
• Task is session-level metadata — start(rootInput, task?) accepts an optional Task = { description: string }. It rides on crew.started once and lives in the snapshot for replay/observability. It does not propagate onto per-step requests; it's not a substitute for the input cascade.
• Snapshot/resume preserves it all — session.snapshot() returns a serializable record of every pending step's input, role snapshot, memex context, the session-level task, and inner-session state. resumeCrewSession(snap, laterEvents, opts) rebuilds it.

system_prompt is optional and unenforced

Each role can declare a system_prompt: string on the shape. It's carried in the RoleSnapshot embedded in agent.step.requested so workers can use it as the model's system message. The library does not validate that LLM-driven roles (capabilities.thinking_allowed: true) actually have one — that's a worker-side / caller-side concern. If you want strict checks, walk the parsed shape yourself before buildCrewGraph and reject roles that need a prompt and don't have one.

Determinism

Everything is deterministic given the inputs:

• Same (shape, counts, ordered inbound events) → byte-identical outbound event log (after canonicalize()).
• Inbound delivery order doesn't affect outbound — votes are tallied in graph declaration order.
• Clock and ids are injected; the library never calls Date.now() or generates random data.
• tick() is caller-driven — timeouts only fire when the caller calls it.

This makes replay testing trivial and snapshot/resume correct by construction.

What's in the box

Concept	Public surface	Source
Schema validation	CrewShapeSchema, CrewAgentsSchema	src/schema/
Graph builder + linter	buildCrewGraph, lintCrewShape	src/graph/
Event vocabulary	CrewEvent discriminated union	src/events/types.ts
Canonical JSON	canonicalize	src/events/canonical.ts
Correlation ids	correlationId, hashHex	src/runtime/ids.ts
Clock injection	systemClock, frozenClock	src/runtime/clock.ts
Hooks	createHookRegistry, HookContext	src/runtime/hooks.ts
Session machine	createCrewSession, resumeCrewSession	src/session/
Voting	resolveVotes	src/voting/resolve.ts
Checkpoints	createInMemoryCheckpointStore	src/checkpoint/
MemEX adapter (peer)	createMemexAdapter	src/memex/ai2070.ts (subpath @ai2070/crew/memex)

See API.md for the full reference.

Status

Every section of PLAN.md is implemented:

• ✅ Schemas + graph builder + linter
• ✅ Event vocabulary + canonical JSON + deterministic correlation ids
• ✅ State machine: start / deliver / tick / cancel
• ✅ Hooks (sync), permissions (ACL gates), fixer activation, nested crews
• ✅ MemEX adapter (memex_context out, memex_commands in, hard isolation)
• ✅ Snapshot / resume (with ResumePolicy) + CheckpointStore
• ✅ Voting: first_valid, majority, unanimous, weighted_consensus (equal weights)

Deferred to v2 (per plan): dynamic crews / permissions.invite, custom voting weight functions, best_of_n voting (needs scoring source), cost tracking.

License

Apache-2.0