Agents

Ask for work “in parallel” and the model calls spawn_agent for each part. Each agent gets its own tools, its own step limit, and its own thinking effort — and runs alongside your chat, which keeps going.

The Agents rail

⇧⌘A opens the Agents rail on the right:

  • one row per agent, with a progress bar and its current step,
  • click a row to open that agent's transcript beside your chat, and read along while the main conversation continues,
  • cancel one, or cancel all — ⌘. stops a reply and, with agents running, doubles as cancel-all.

Results come back into the main reply. Nothing in the chat is blocked while agents work.

Your caps

Settings → Agents:

  • Concurrency — how many run at once (default 3). Splash works on several requests together, and how much that helps depends on what the agents are doing:
    • Writing gains a lot. On a 32 GB Mac mini with a 27B model, four short answers at once each came out about as fast as one alone, so the four finished in roughly the time of one.
    • Reading does not, unless they read the same thing. Four agents each given a different long document to read first (about 9,000 tokens) took as long together as one after another. Four given the same opening read it once between them and finished in the time of one, which is why every agent is sent the same instructions and tools before its own task.
    • Each agent in flight also holds its own context in memory, so on a 32 GB server three or four is a sensible ceiling. If replies slow to a crawl or an engine restarts, lower it.
  • Step limit — per agent (default 25). An agent that has not converged by then is stopped and reported as such, rather than running to the end of the window.
  • Thinking effort — per agent, independent of the main chat. Cheap and fast by default; raise it for the one agent doing the hard part.

What they are good at

  • Research fan-out — one agent per source or question, merged in the main thread.
  • Codebase surveys — “every call site of X, every test covering Y, every TODO in Z”, three agents, one answer.
  • Drafting in parallel — three variants of a deliverable, chosen by you afterwards.

They are not more intelligence — they are more time. Each agent uses the same model with the same limits; agents buy wall-clock time and parallel context, not better answers.

Found a problem, or is something here unclear? Report it.