Subagents

Subagents let a chat delegate a bounded task to another model — local or remote — or to one of your saved agents, and get a compact result back without cluttering the conversation. Offload research to a bigger cloud model, hand a coding question to a specialist agent, or generate an image, all mid-turn.

Delegation is off by default, approved on first use, and configured per agent.

Spawn tools

Tool	What it does
spawn_agent(input, agent)	Delegate a task to one of your saved agents — it uses that agent's persona and allowed child tools
spawn_model(input, model)	Delegate to a bare model (no persona) — useful for "ask a bigger/faster model" moments
spawn_batch(jobs)	Fan out independent jobs across a mixed set of allowed agents and bare models

The subagent runs its own bounded job and returns a single compact result. Its inner steps render live in the chat row but never enter your conversation's transcript, so context stays lean.

For spawn_batch, every job carries a stable ID, an agent or model target, and its own input. One batch can mix saved agents, local models, and remote models. Osaurus validates the whole batch first, asks for one approval for the batch, and returns one result per job in the same order as the input.

Three more capabilities share the same machinery and the same per-agent settings surface:

• image — generate or edit a picture inline; see Image Generation
• computer_use — drive a macOS app; see Computer Use
• browser_use — drive a persistent, isolated browser; see Browser Use

Getting started

1. Open Agents → Abilities → Subagents on the agent you want to grant delegation (the main chat has its own Subagents tab too).
2. Enable Spawn & Delegation and pick the spawnable pool — which saved agents and models this agent may delegate to. The pool is searchable, and you can attach a note to each entry to tell the model when to use it.
3. Ask for something that benefits from delegation: "Have the research agent summarize this paper, then continue." For independent work, ask it to run the tasks as a batch.
4. Approve the first use. You can set Spawn to Always Allow in the same settings surface.

Batch execution and residency

Delegation works in any direction — local to local, local to remote, remote to local, remote to remote. Only one case touches GPU memory: delegating from a local model to a different local model.

Direction	Behavior
Local → same local model	Runs in place — no swap
Local → different local model (handoff on)	Unload the chat model, run the job, reload, continue
Local → different local model (handoff off)	Rejected up front — nothing is evicted
Local ↔ remote, remote ↔ remote	Runs in place

Local Orchestrator Handoff is on by default. With handoff on, two large models never fight for memory; with it off, Osaurus refuses the delegation cleanly instead of erroring mid-run.

spawn_batch groups work by resolved model:

• Jobs using the same local model reuse one loaded model and can run concurrently through continuous batching.
• Jobs using different local models run as serial waves, so cold loads and residency handoffs never race.
• Remote jobs start independently and can overlap every local wave.
• Before each local wave, RAM admission considers the model footprint, per-child working memory, current engine occupancy, and continuous-batching capacity. A wave may be narrowed, split into subwaves, or refused before unloading a model.

Results still preserve the original job order, regardless of completion order.

One concurrency setting

Max subagents per batch and Server → Concurrent Sessions are the same canonical setting (1–32). Changing either updates the other. It is a ceiling for accepted batch size and local fan-out; RAM admission, engine occupancy, continuous batching, and local-model grouping can reduce the concurrency of a particular wave.

Per-agent configuration

Everything is scoped to the agent, in its Subagents tab:

• Spawn & Delegation — enable, plus separate allow-lists of spawnable agents and bare local or remote models
• Permission — Ask, Always Allow, or Deny; Ask produces one prompt for an entire spawn_batch
• Child tools — choose None or Read Only. A saved-agent child can receive the cancellation-audited subset of its enabled tools; a bare-model child has no target-agent tools. Read Only can additionally grant bounded file_read / file_search access.
• Budgets — delegate tokens, turns, tool calls, elapsed time, and the shared batch/concurrency limit
• Image — enable the image tool and pick the agent's image model
• Computer Use — enable, autonomy ceiling, screen context (details)
• Browser Use — enable, plus an optional model override (details)
• AppleScript — enable the AppleScript subagent and pick its model

Spawn remains unavailable until its allow-list is configured; the other delegation capabilities ship disabled.

---

Related:

• Agents — creating and configuring agents
• Image Generation — the image subagent
• Computer Use — the computer_use subagent
• Browser Use — the browser_use subagent
• Models — local and remote models a subagent can target