ORBIS

Appearance

The agent model ​

ORBIS's brain is deliberately a small, fast model that routes — not a giant model that does everything. This page explains that choice and how delegation, orchestration, and tools fit around it.

For the knobs, see the Agent reference.

Router first ​

Every turn, the LLM makes a routing decision before anything else:

  • Answer directly — most conversation.
  • Call a tool — a concrete action (set a reminder, list reminders).
  • Delegate — hand a heavier task to a specialist agent.

Keeping the routing model small and fast is what makes the orb feel responsive. The heavy thinking, when it's needed, happens elsewhere — in a delegate — and the orb narrates the result. (If you want a stronger model for just the decision turn, that's what two-model routing is for.)

Delegates: specialists the orb calls ​

A delegate is another agent ORBIS can hand work to. You give each one a name and a description; the LLM reads the descriptions and picks the right delegate for the task. Two kinds:

  • A2A agents speak the Agent2Agent protocol — a standard JSON-RPC contract for agents to message each other, keep a conversation across turns, and stream progress. This is how ORBIS talks to other agents in a fleet.
  • OpenAI-compatible delegates are any /v1/chat/completions endpoint treated as a sub-agent (a larger model, a specialized service).

The point: the orb stays a fast conductor; the delegates do the deep work.

Sync vs. background ​

Delegation comes in two shapes:

  • Delegate (sync-ish). The orb hands off one task, waits, and speaks the answer. It acknowledges immediately ("on it…") so it isn't silent, and it can narrate progress a delegate streams back.
  • Orchestrate (background). For a multi-step ask ("get the fleet status, then dig into any incident"), the orb runs a bounded loop in the background — calling delegates step by step, keeping context with each one — and reports the synthesized result when it's done. You can keep talking meanwhile.

Both are designed so a slow or unreachable delegate can't hang the orb: every step is time-bounded, and a dropped connection falls back gracefully.

Reminders and other actions ​

Actions the orb can take itself (not delegated) are tools — small, exact capabilities. Reminders are the headline example: set, list, and cancel them by voice, and see them in the reminders bell. A tool result is fed back to the model so it knows the action succeeded and can confirm it to you.

Why this shape ​

  • Responsiveness — a small router replies fast; depth is opt-in.
  • Composability — any A2A or OpenAI-compatible agent becomes a delegate without changing ORBIS.
  • Resilience — time-bounds and failover keep the orb talking even when a model or delegate misbehaves.

See also ​