ORBIS

Appearance

Add a delegate ​

A delegate is a sub-agent ORBIS can hand work to. This guide adds one and confirms the orb can reach it. For the concepts, see The agent model; for every field, the Agent reference.

Before you start ​

Decide which kind you're adding:

  • A2A agent — another agent that speaks the Agent2Agent JSON-RPC protocol (e.g. a peer in your fleet). You'll need its endpoint URL.
  • OpenAI-compat — any /v1/chat/completions endpoint you want to treat as a specialist (a bigger model, a service).
  • ACP coding agent — a terminal coding agent (protoCLI, OpenCode, Claude Code, Codex) ORBIS launches and drives by voice. It has its own guide: Voice-drive a coding agent.

If the endpoint needs a secret, put it in your .env on the host first — the UI references the env-var name, never the value.

Steps ​

  1. Open Settings → Agent → Delegates and click Add delegate.
  2. Pick the type — A2A agent or OpenAI-compat.
  3. Fill in the shared fields:
    • Name — lowercase, no spaces (e.g. helm). This is what the model uses in delegate_to(target=…).
    • Description — be specific. The model reads this to decide when to pick this delegate over another. "Chief of staff — sitreps, planning, fleet delegation" beats "an agent".
    • URL — for A2A, the JSON-RPC endpoint (often ending in /a2a); for OpenAI-compat, the base URL ending in /v1.
  4. Fill in the type-specific fields:
    • A2A — pick an Auth scheme (apiKey, bearer, or none) and, if used, the Credentials env var holding the secret.
    • OpenAI-compat — the Model id and, if the endpoint needs auth, the API key env var. Optionally a system prompt.
  5. Click Test — ORBIS pings the endpoint and reports reachability and latency before you save, so you can fix a typo without committing it.
  6. Click Create.

Use it ​

Just talk: ask the orb something that fits the delegate's description — e.g. "ask helm for the project status." The orb acknowledges, hands off, speaks the result. For multi-step asks, it can orchestrate across delegates in the background.

Troubleshooting ​

  • Test fails — re-check the URL (A2A endpoints usually end in /a2a, OpenAI in /v1) and that the credentials env var is set in .env on the host.
  • Marked "⚠ unconfigured" in the list — the runtime rejected the entry; fix the config or it won't be offered to the model.
  • The orb never picks it — sharpen the description; that's the only thing the model uses to choose.

See also ​