Skip to main content

Context

Context is everything the agent can see while it works — in LLM terms, the context window: the full set of messages sent to the model on each turn.

This matters because most "the agent is being dumb" moments are really context problems. An agent that gives vague answers usually doesn't have the information it needs. An agent that ignores instructions often has so much in context that the important part got buried. Once you know what goes in, you can fix both.

What the agent sees​

On every turn, Taurus assembles the context from:

  • Platform instructions — Taurus's own system prompt, telling the agent about its environment, tools, and platform conventions. Always present; you don't edit this.
  • Your agent prompt — the system prompt you wrote in the agent's settings. Also always present, and never compacted — which makes it the most reliable place for anything the agent must never forget.
  • Memory — the top of /workspace/MEMORY.md (roughly the first 16 KB; the exact limit depends on the model) is injected at the start of every run. The rest of the file stays on disk — the agent can read it with a tool call, but it isn't automatic. This is why agents keep the important stuff at the top and push detail into separate files.
  • Team roster — if the agent has children, their names and what each one does, so it knows whom it can delegate to.
  • Targets — any remote machines the agent has been granted access to.
  • The conversation so far — every message, tool call, and tool result in the current run. By far the largest part, and it grows with every turn. This is why long runs eventually hit the wall.

Messages you send while the agent is mid-work are inserted right after the current tool call finishes, so the agent sees them on its next step.

Running out of room​

Every model has a fixed context window — 200K tokens for Claude, for example. The context ring in the run footer shows how full it is at a glance: gray is fine, amber (60–85%) means compact soon, red means compact now.

Compaction​

Compaction summarizes older messages into a condensed recap so the run can keep going. The summary replaces the original messages — the key facts survive, the fine detail doesn't.

  • /compact (or /compact full) — summarizes everything except the last few messages. Frees the most room.
  • /compact half — keeps roughly the recent half of the conversation verbatim. Gentler, frees less.

If a running agent hits its limit, auto-compaction kicks in on its own so the agent can keep working. But it's better to compact proactively — when the ring turns amber, /compact half lets you pick the moment and keep more of the recent detail. You can also just click the context ring.

Working with context, not against it​

  • Continue the run when the next message builds on the last one. The agent keeps everything it learned.
  • Start a new run for unrelated work. Old history is pure noise for a new task.
  • Fork to explore alternatives. Each branch gets its own context from the fork point onward.
  • Split huge tasks. A 200-turn task will compact several times and shed detail each time. Break it into focused runs — or let the agent push chunks into subruns or child agents, each with a fresh window.