Multi-Agent Teams

Some questions are too broad for one agent to research in one context window. Teams is agentFast's own native multi-agent orchestrator — a supervisor decides what to investigate, sub-agents each take one angle as a full run of their own, and their real cost rolls up into one number on the parent. It is registered through the same adapter seam as LangGraph, CrewAI, and the rest — the sixth orchestration option, not a bolt-on.

How a fan-out run actually works
supervisor plans anglesN sub-agents run in parallelsynthesis merges findingsone run, one cost
Each child is a full run, not a step

Its own context window

One angle's research can't crowd out another's or push the parent toward compaction — each child gets a full window of its own.

Its own budget

worker_max_iterations and worker_max_tokens are sized per child, not shared with siblings — one runaway angle can't starve the others.

Its own step tree

Every LLM call, tool call and guardrail check a child makes is a full trace, exactly like a single-agent run — nested under the parent in the dashboard.

Its own crash recovery

A child checkpoints exactly the way any other run does. The roster is written before any child starts, so resume never loses or re-runs finished work.

What makes the cost trustworthy
  1. 1
    Admission is lazy, not upfront

    Children are admitted up to max_parallel as slots free — not all planned at once — because a child's usage only reaches the parent once it actually finishes.

  2. 2
    Real numbers, read back from the store

    The moment a child finishes, its persisted token counts and cost — not the in-memory result — are added to the parent's totals.

  3. 3
    The ceiling sees real accrued spend

    Because admission tops up as children complete, the parent's budget check is judged against what the team has actually spent, not an upfront guess.

bash
# demo/research.sh — driven directly against the runtime
sdk: team
team:
  max_parallel: 2
  max_tasks: 2
  worker_max_tokens: 40000
"One run id, one number for what the team cost"
Every child's cost rolls up into the parent the instant it finishes — you never add up children by hand, and the dashboard nests each one under its parent as its own expandable step tree. That is what makes a multi-agent system billable and governable instead of a science project.
Honest limits

Flat, not recursive

Children cannot spawn grandchildren — the supervisor plans a flat list once, up front. That keeps the budget ceiling and recursion depth both bounded.

Children don't talk to each other

Each child is independent by construction. If one angle's findings should inform another, that happens in synthesis, after both finish.

A dead angle doesn't sink the team

on_child_failure: continue (the default) synthesises from whatever succeeded, and names the angle that didn't in the report.

Teams shares the exact same durable execution, guardrails, and observability every other adapter gets — see it running on the ResearchAgent, or read the full mechanics in the docs.

Stop rebuilding the plumbing.

agentFast is the production layer — memory, 50 tools, observability, evals, guardrails, human-in-the-loop — for LangGraph, CrewAI, the Claude & OpenAI Agent SDKs & Vanilla. Own it for life.

Get agentFast — from $89