autosearch-hitl

autosearch-hitl

Autonomous optimization loop. Generalizes the autoresearch pattern ("change → measure → keep/discard") to any viable domain.

How to use it (plain language)

You don't need to be technical. Just tell the agent what you want to make better — even casually:

• "how can I improve this?"
• "this feels slow — make it faster"
• "make my tests cover more"
• "can this be cheaper to run?"
• "make these search results better"

You don't need to know how to measure success: if it's unclear, the skill asks you and helps set up a measurement first. It then works on an isolated copy, keeps changes that improve the number and reverts the rest, and stops when it stops getting better — so your real code stays safe.

---

The steps below are the internal procedure the agent follows — you don't need to read them to use the skill.

Step 1 — Understand the goal

Find out (from the prompt or by asking): what should improve and in which direction (lower-is-better or higher-is-better). Identify the domain dynamically (code, prompt, config, data, hyperparameters, etc.) — there is no fixed list.

Step 2 — Check the 4 pre-conditions

The loop is only honest and safe if ALL are satisfied (see caveats below):

1. Controllable artifact — is there something concrete to change?
2. Objective metric — is there a number that says better/worse, with a direction?

   Note: when the domain already has a known measurement harness (e.g. LLM training in autoresearch, whose val_bpb comes from the harness), this pre-condition is considered satisfied by the harness — the user doesn't need to declare it.

3. Repeatable measurement — is there (or can you create) a command that produces that number? This pre-condition is satisfied if the measurement already exists or can be created; it only fails when measuring the goal is genuinely impossible. A missing-but-creatable measurement does not fail — proceed to the general engine (which helps create it in Phase 2).
4. Reversible change — can you undo a bad attempt?