Agents configuration

AgentFlow does not talk to agent APIs directly. It starts each agent as a subprocess without a shell (application/internal/agent/exec), passing a plain argv built from your config. Every key under agents: is a logical id you reference from flags (--agent), from work.default_agent, and from routing rules.

The block below is a realistic shape: several ids point at different executables, with optional args, models, timeouts, and an Ollama-style HTTP endpoint where that agent type expects it.

Example

agents:
  kiro:
    command: kiro
    args: ["--cli"]
  cursor:
    command: cursor-agent
    default_model: auto
    timeout: 3600
  codex:
    command: codex
    timeout: 3600
  claude:
    command: claude
    timeout: 3600
  ollama:
    endpoint: http://localhost:11434
    model: qwen2.5-coder:14b
    embedding_model: nomic-embed-text
    timeout: 300

Fields

The table is a quick map from config field to how the exec layer uses it. Nothing here is implicit: if you omit timeout, defaults elsewhere apply; Ollama-specific keys only matter for an agent that uses that protocol.

Defaults in work

The work block chooses which logical ids run when a command does not override them. Adjust these when your default implementation, reviewer, or enricher should change repo-wide.

work:
  default_agent: cursor
  default_reviewer: codex
  default_enricher: ollama

You can still override per invocation, for example agentflow dev feat --agent cursor or agentflow review feat --agent codex.

Dry-run

With --dry-run, the exec layer does not spawn real subprocesses. That is useful in CI when you only want to validate wiring, and for rehearsing docs or scripts that would otherwise wait on long-running agents.

Related

• Agents overview
• Ollama
• CLI: dev