Crews — Multi-Agent Collaboration

A Crew is a team of agents working together toward a shared goal. One agent decomposes the work, others execute their specialisations, and a reviewer validates the results before synthesis.

Scenario: A PR agency uses a 3-agent crew to produce press releases. The Coordinator decomposes the brief, a Copywriter drafts the release, a Fact Checker reviews claims, and the Coordinator synthesises the final approved version.

Crew member roles

Role	Responsibility
coordinator	Decomposes the goal into tasks and synthesises the final result. Every crew has exactly one coordinator.
worker	Executes assigned tasks. The workhorse of the crew.
qa	Validates output quality. Can request revisions before the coordinator synthesises.

Per-member permission policies

Each crew member can be given an individual policy that constrains what the agent is allowed to do during execution. Policies are stored in the member's config JSONB field and enforced by CrewOrchestrator at runtime.

Field	Description
tool_allowlist	Array of tool IDs the agent may use. If set, the agent cannot call tools outside this list.
max_steps	Maximum number of execution steps allowed for this member in a single crew run.
max_credits	Credit budget cap for this member. Execution stops when the member's spend reaches this limit.

json

// POST /api/v1/crews  (or PUT /api/v1/crews/CREW_ID)
{
  "name": "PR Agency Crew",
  "members": [
    {
      "agent_id": "agt_01jq...",
      "role": "worker",
      "workerConstraints": {
        "tool_allowlist": ["tool_web_search", "tool_fetch_url"],
        "max_steps": 20,
        "max_credits": 500
      }
    }
  ]
}

Process types

Sequential

Agents execute one at a time. Each agent receives the previous agent's output. Predictable, auditable, but slower.

Parallel

All agents execute simultaneously on independent sub-tasks. Fastest option. Results merged by the coordinator.

Hierarchical

Tree structure. The coordinator assigns sub-tasks to workers who may delegate further. Best for complex, decomposable goals.

Running a crew

From the crew detail page, click Execute or navigate to /crews/{id}/execute. Provide a goal and optional context. The crew starts immediately.

bash

curl -X POST https://fleetq.169.58.89.204.sslip.io/api/v1/crews/CREW_ID/execute \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -d '{
    "goal": "Write a press release for our Series B funding announcement",
    "context": "Amount: $12M. Lead investor: Acme Ventures. Use case: AI for e-commerce."
  }'

Under the hood: CrewOrchestrator

The CrewOrchestrator runs a 4-step loop:

DecomposeGoal

The coordinator agent breaks the goal into discrete tasks assigned to each crew member.

Execute tasks

Workers execute their assigned tasks (in parallel or sequentially, based on process type).

ValidateTaskOutput

The QA agent validates each output. Failed validations trigger a retry of that specific task.

SynthesizeResult

The coordinator merges all validated outputs into the final result and saves it as an Artifact.

Crew artifacts

Each crew execution automatically collects artifacts from task outputs via CollectCrewArtifactsAction. View and download artifacts from the execution detail page.

MCP tools

Tool	Purpose
crew_list	List crews with filtering
crew_get	Get crew details including members
crew_create	Create a new crew with members
crew_update	Update crew configuration
crew_execute	Start a crew execution
crew_execution_status	Check execution progress
crew_executions_list	List past executions

API endpoints

Method	Path	Purpose
GET	/api/v1/crews	List crews
GET	/api/v1/crews/{id}	Get crew details
POST	/api/v1/crews	Create crew
PUT	/api/v1/crews/{id}	Update crew
DELETE	/api/v1/crews/{id}	Delete crew
POST	/api/v1/crews/{id}/execute	Start execution
GET	/api/v1/crews/{id}/executions	List executions
GET	/api/v1/crews/{id}/executions/{eid}	Execution detail

See also: Agents (individual AI workers), Workflows (visual DAG pipelines).