Examples

Ready-to-use orchestration configs for common team structures.

---

Software development team (default)

Four-agent pipeline: Planner → Developer → Tester → Reviewer. Validates handoffs with routing validators and disk-based artifacts.

A complete version of this config is in config/examples/orchestration.yaml.

Key features: - Keyword routing with RequireBrief, RequireWriteFile, RequireShellPass, and TestReportValid validators - SourceAgents restrictions on every route — agents can only trigger transitions intended for their role - RequiredCommandPattern on the HANDOFF TO TESTER shell-pass route (go build|go test) — prevents satisfying the check with unrelated commands - brief.json and test-report.json written to disk to survive context compaction - Compaction enabled (trigger at 30 turns, keep 8); post-compaction note prompts agents to re-read artifacts - Stuck detection throws after 3 consecutive bad turns (no keyword, wrong-role keyword, multi-keyword, or validator failure) - xAI Grok models via the Models registry - USD spending cap

---

Single agent (minimal)

A single agent with filesystem access — good for quick tasks or as a starting point.

Orchestration:
  Name: Assistant
  Agents:
    - Name: Assistant
      Instructions: >-
        You are a helpful software assistant. Use the tools available to read and write
        files, run shell commands, and complete the user's task. When you are done,
        say TASK COMPLETE.
      Model:
        ModelId: gpt-4o
      Plugins:
        - FileSystem
        - Shell
        - Search
      FunctionChoice: auto
  Selection:
    Type: sequential
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: TASK COMPLETE
      - Type: maxiterations
        MaxIterations: 20
  MaxCostUsd: 1.0

---

Research team

Two agents: a Researcher who gathers information via web requests and a Writer who synthesises it into a document.

Orchestration:
  Name: ResearchTeam
  Agents:
    - Name: Researcher
      Description: Gathers information from the web and structures findings.
      Instructions: >-
        You are a research assistant. Use http_get to fetch content from provided URLs
        or APIs. Summarise your findings in structured form. When you have gathered
        enough information, write HANDOFF TO WRITER.
      Model:
        ModelId: gpt-4o
      Plugins:
        - Http
        - Json
        - FileSystem
      FunctionChoice: auto
    - Name: Writer
      Description: Synthesises research into a polished document.
      Instructions: >-
        You are a technical writer. The Researcher has gathered information for you.
        Write a clear, well-structured document based on their findings. Save the final
        document to disk using write_file. When done, say DOCUMENT COMPLETE.
      Model:
        ModelId: gpt-4o
      Plugins:
        - FileSystem
      FunctionChoice: auto
  Selection:
    Type: keyword
    DefaultAgent: Researcher
    Routes:
      - Keyword: "HANDOFF TO WRITER"
        Agent: Writer
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: DOCUMENT COMPLETE
        AgentNames:
          - Writer
      - Type: maxiterations
        MaxIterations: 15
  MaxCostUsd: 1.50

---

Local model team (Ollama)

Two agents running entirely on local models via Ollama. No API keys required.

Orchestration:
  Name: LocalTeam
  Models:
    local-fast:
      ModelId: llama3.2
      MaxTokens: 4096
    local-coder:
      ModelId: qwen2.5-coder:7b
      MaxTokens: 8192
  Agents:
    - Name: Planner
      Instructions: >-
        You are a planner. Analyse the task and write a clear plan with numbered steps.
        When done, write HANDOFF TO DEVELOPER.
      Model:
        ModelId: local-fast
      FunctionChoice: none
    - Name: Developer
      Instructions: >-
        You are a developer. Follow the plan provided and implement the changes using
        the tools available. When done, say TASK COMPLETE.
      Model:
        ModelId: local-coder
      Plugins:
        - FileSystem
        - Shell
      FunctionChoice: auto
  Selection:
    Type: keyword
    DefaultAgent: Planner
    Routes:
      - Keyword: "HANDOFF TO DEVELOPER"
        Agent: Developer
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: TASK COMPLETE
      - Type: maxiterations
        MaxIterations: 20

---

Mixed provider team

Different agents use different providers: GPT-4o for planning, Claude for implementation, a local Llama for review.

Orchestration:
  Name: MixedProviderTeam
  Models:
    planner-model:
      ModelId: gpt-4o
      MaxTokens: 4096
    coder-model:
      ModelId: claude-3-5-sonnet-20241022
      MaxTokens: 16384
    reviewer-model:
      ModelId: llama3.2
      MaxTokens: 4096
  Agents:
    - Name: Planner
      Instructions: Analyse the task and produce a structured plan. End with HANDOFF TO DEVELOPER.
      Model:
        ModelId: planner-model
      Plugins:
        - FileSystem
        - Search
      FunctionChoice: auto
    - Name: Developer
      Instructions: Implement the plan using the tools available. End with HANDOFF TO REVIEWER.
      Model:
        ModelId: coder-model
      Plugins:
        - FileSystem
        - Shell
        - Git
      FunctionChoice: required
    - Name: Reviewer
      Instructions: >-
        Review the implementation. If it meets the requirements, say APPROVED.
        If not, say CHANGES REQUIRED and describe what to fix.
      Model:
        ModelId: reviewer-model
      Plugins:
        - FileSystem
      FunctionChoice: auto
  Selection:
    Type: keyword
    DefaultAgent: Planner
    Routes:
      - Keyword: "HANDOFF TO DEVELOPER"
        Agent: Developer
      - Keyword: "HANDOFF TO REVIEWER"
        Agent: Reviewer
      - Keyword: "CHANGES REQUIRED"
        Agent: Developer
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: \bAPPROVED\b
        AgentNames:
          - Reviewer
      - Type: maxiterations
        MaxIterations: 30
  MaxCostUsd: 2.0

---

Magentic team

A manager LLM plans the work and coordinates a Researcher and Developer dynamically each round. No routing keywords or conditions needed — the manager reasons about what has been done and picks the right agent next.

A complete version of this config is in config/examples/magentic-team.yaml.

Key features: - Manager LLM drives a two-level loop: orientation (once) then per-round ledger evaluation + speaker selection - Participant agents need no special routing keywords — they just do the work - Built-in stall detection: if MaxStallCount consecutive rounds make no progress, the manager replans - Termination section is ignored; session end is controlled by MaxRoundCount, MaxStallCount, and MaxResetCount - Named model aliases let you assign a stronger reasoning model to the manager and a faster model to workers

Orchestration:
  Name: Magentic Team
  Description: >
    AI-managed team orchestrated by Magentic. A manager LLM plans the work,
    dynamically selects participants each round, and replans if progress stalls.

  Models:
    manager:
      ModelId: gpt-4o          # upgrade to o3 or claude-opus-4-6 for complex tasks
    worker:
      ModelId: gpt-4o-mini

  Agents:
    - Name: Researcher
      Description: Gathers information, searches, and produces sourced summaries.
      Instructions: |
        You are a Researcher. Find information, analyse it, and produce well-sourced
        summaries. Use your tools to search and read content. Be thorough but concise.
      Model:
        ModelId: worker
      Plugins:
        - FileSystem
        - Search
        - Scratchpad

    - Name: Developer
      Description: Writes code, implements features, runs tests, and fixes bugs.
      Instructions: |
        You are a Developer. Write clean, working code that solves the problem.
        Implement what is asked, verify with shell_run, and report results accurately.
      Model:
        ModelId: worker
      Plugins:
        - FileSystem
        - Shell
        - Git
        - Scratchpad

  Selection:
    Type: magentic
    Magentic:
      Model:
        ModelId: manager
      MaxRoundCount: 20
      MaxStallCount: 3
      MaxResetCount: 2
      EnablePlanReview: false

  # NOTE: Termination is ignored for Selection.Type 'magentic'.
  Termination:
    Type: maxiterations
    MaxIterations: 50

  Compaction:
    TriggerTurnCount: 50
    KeepRecentTurns: 10

  Checkpoint:
    Mode: json
    Path: .fuseraft/checkpoints

Generate this config with fuseraft init --template magentic.

---

With MCP tools

Adding an MCP server that provides browser automation tools alongside built-in plugins:

Orchestration:
  Name: WebAgent
  McpServers:
    - Name: Browser
      Transport: stdio
      Command: npx
      Args:
        - "-y"
        - "@modelcontextprotocol/server-puppeteer"
  Agents:
    - Name: WebScraper
      Instructions: >-
        You are a web scraping agent. Use the browser tools to navigate pages, extract
        data, and save it to files. When done, say TASK COMPLETE.
      Model:
        ModelId: gpt-4o
      Plugins:
        - Browser
        - FileSystem
        - Json
      FunctionChoice: auto
  Selection:
    Type: sequential
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: TASK COMPLETE
      - Type: maxiterations
        MaxIterations: 15

---

With sandbox and cost cap

A config safe to run on untrusted tasks — file access restricted to a specific project, HTTP limited to known hosts, and a hard cost cap:

Orchestration:
  Name: SafeTeam
  Agents: [ ... ]
  Selection: { ... }
  Termination: { ... }
  Security:
    FileSystemSandboxPath: /home/user/projects/sandboxed-project
    HttpAllowedHosts:
      - api.github.com
      - pypi.org
  MaxCostUsd: 0.50

---

With compaction for long tasks

For tasks that may run for many turns (e.g. large refactors, multi-file migrations):

Orchestration:
  Name: LongRunningTeam
  Compaction:
    TriggerTurnCount: 40
    KeepRecentTurns: 10
    Model:
      ModelId: gpt-4o-mini
  Agents: [ ... ]
  Selection: { ... }
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: \bAPPROVED\b
      - Type: maxiterations
        MaxIterations: 100

When 40 messages have accumulated, the orchestrator summarises the oldest 30 turns into a single message and continues with the 10 most recent turns verbatim. This can run indefinitely without hitting context limits.

---

Structured routing — content review pipeline

Three agents use JSON output instead of routing keywords. The Reviewer returns {"review_result": "Yes"} or {"review_result": "No"} and the orchestrator routes based on that field value, with no keyword scanning involved.

Orchestration:
  Name: ContentPipeline
  Agents:
    - Name: Drafter
      Instructions: >-
        Write a draft based on the user's request. Return your response as a JSON
        object: {"draft_content": "<your draft here>"}. Your entire response must
        be valid JSON.
      Model:
        ModelId: gpt-4o
      FunctionChoice: none
    - Name: Reviewer
      Instructions: >-
        Review the draft in the conversation. Check that it is accurate and
        well-written. Return your decision as a JSON object:
        {"review_result": "Yes", "reason": "<explanation>"} if acceptable, or
        {"review_result": "No", "reason": "<what to fix>"} if it needs revision.
        Your entire response must be valid JSON.
      Model:
        ModelId: gpt-4o
      FunctionChoice: none
    - Name: Publisher
      Instructions: >-
        The draft has been approved. Format it as a final polished document and
        print it. End your response with the word PUBLISHED on its own line.
      Model:
        ModelId: gpt-4o
      FunctionChoice: none
  Selection:
    Type: structured
    DefaultAgent: Drafter
    StructuredRoutes:
      - Agent: Reviewer
        Condition:
          Field: draft_content
          Exists: true
        SourceAgents:
          - Drafter
      - Agent: Publisher
        Condition:
          Field: review_result
          Is: "Yes"
        SourceAgents:
          - Reviewer
      - Agent: Drafter
        Condition:
          Field: review_result
          Is: "No"
        SourceAgents:
          - Reviewer
  Termination:
    Type: composite
    Strategies:
      - Type: regex
        Pattern: PUBLISHED
        AgentNames:
          - Publisher
      - Type: maxiterations
        MaxIterations: 15
  MaxCostUsd: 1.0

Key differences from keyword routing: - Agents return JSON objects — routing happens on field values, not on specific text appearing in the response - No APPROVED/REVISION REQUIRED style keywords needed - Session end is handled by the Termination regex strategy rather than a self-routing route - The --diagram flag renders conditions as edge labels: review_result = Yes, draft_content exists