gutDocs Get started

For developers building agents

Agents think too much. Give them a gut.

Let your agent’s big model think once, to set the goal. gut hands every small pick after that to a fast decision model, round by round, until the goal is met.

How gut plays a game: the wiki race

  1. Imagine a WikiRace from Banana to Roman Empire.

    Get from one Wikipedia page to the other using only its links.

  2. You land on Banana. To play, you need three things.

  3. A goal: where you’re going.

  4. Context: where you are now.

  5. Ops: what you can do. Here, click a link.

  6. Then you decide.

  7. Now swap yourself for an AI.

  8. With gut, those three things are the whole task.

  9. Run it. You wrote the moves; gut finds the path.

You know the moves. You don’t know the path.

gut is for paths you can’t write down, made of steps that are easy: finishing a checkout on a site you haven’t seen, or finding a page in docs you don’t know. In a browser, @gut.run/playwright turns a page’s links, buttons and fields into ops.

You know…

  1. the steps and their orderUse plain code
  2. the steps and their order, with a judgment at fixed points ("is this a billing ticket?")Use plain code that asks a decision model at those points
  3. every move that is allowed, but not the order: it depends on what the world shows each timeUse gut
  4. not even the moves, or each step needs real reasoning or writingUse a full LLM agent

The wiki race in 39 lines. Without gut, 149.

with gut

without

Each bar is a line. Without gut, the race spells out the bundles, the split on refusal, the budget and the log: wikirace.gut.ts against wikirace.plain.ts.

Quick enough for every small step.

255 ms

The median request to Jev, a hosted decision model. It only picks; it never writes text.

How it was measured, and the browser bench

See why it picked what it picked.

The race above, replayed in gut’s inspector: 14 rounds, 27 requests and 40,209 input tokens to clef-flash on a laptop. Open any turn to read what the model read, and the probability it gave each option.

Get started

Try a task with no model first, answering each pick yourself in the inspector. Then point it at Jev (hosted, through OpenRouter), Clef on a local Ollama, or any /v1/systemone endpoint.

npm install @gut.run/core && npm install --global @gut.run/cli