Profiles & Delegation

Profiles

Profiles are named agent configurations defined under profiles: in config. Each profile can override:

model — use a different LLM
instructions — custom system prompt
tools — allowlist of available tools
temperature — LLM temperature
maxToolRounds — max tool call iterations
hooks — beforeRun/afterRun tool calls

yaml

profiles:
  researcher:
    instructions: "Search the web and summarize findings."
    tools: ["web_search", "web_fetch", "memory"]
    temperature: 0.5
    maxToolRounds: 8

  coder:
    model: "qwen3-coder:30b"
    instructions: "You are a code assistant."
    tools: ["exec", "read", "write", "memory"]
    maxToolRounds: 15
    hooks:
      afterRun:
        - tool: memory
          args: { action: "append", file: "work-log.md", content: "{{response}}" }

Using profiles

From the CLI:

bash

pnpm run start -- --profile researcher --message "Find AI news"

From cron jobs:

yaml

cron:
  jobs:
    - name: "daily-research"
      schedule: "0 9 * * *"
      prompt: "Research today's AI news"
      profile: "researcher"

Profile resolution

resolveProfile() merges a named profile with the agent defaults. The profile's tool list acts as an allowlist — only the named tools (from the full set) are made available. Other settings fall back to the agent-level defaults.

Delegation

The delegate tool lets the agent spawn a sub-agent with a specific profile. Sub-agents are depth-1 only (they don't get the delegate tool themselves).

Each delegation creates an ephemeral session keyed delegate:<parentSessionId>:<uuid>.

Synchronous delegation

By default, the delegate tool blocks until the sub-agent completes:

delegate({ profile: "researcher", message: "Find the latest AI papers" })

Async delegation

Set async: true to fire the sub-agent in the background and get a task ID back immediately:

delegate({ profile: "researcher", message: "Research topic X", async: true })
→ "Task task_abc123 started"

Background tasks

Tasks are tracked in-memory (intentionally ephemeral — they don't survive restarts). The task_status tool lets the agent list all tasks or check on a specific one:

task_status({ action: "list" })
task_status({ action: "get", task_id: "task_abc123" })

Tasks track status (running, completed, failed), timing, and result/error.