> ## Documentation Index
> Fetch the complete documentation index at: https://docs.withwhile.com/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Install with `uv add whileai`; import as `import whileai as wai`.
> Run the offline path first (`simulator=False`, `wai.seeded_agent`, a callable judge); no key is needed for it.
> Report every pass rate with its interval and n, as `scored.pass_at` prints it.

# Name things after your work

> The platform shows the names your code sends. A coding agent names the agent, behaviors, versions and experiments in the team's own words, and the test by its content.

Every name on [withwhile.com/platform/runs](https://withwhile.com/platform/runs) comes from your code, and the page marks it so. The coding agent that writes the `track(...)` calls decides what the team reads for months. One rule covers it: **read the repo, then use its words.** The package name, the prompt files, the policy doc, the existing test names, the deploy tags and the changelog already hold the vocabulary. The SDK's example names (`refund-agent`, `refund_policy`, `v1`) are for the playbooks, not for a real repo.

| Thing                                       | Rule                                                                                                                 | From the repo                                   | Not this                          |
| ------------------------------------------- | -------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------- | --------------------------------- |
| Agent, `track("…")`                         | the product or service the agent is, kebab-case, one per deployed system; the model is a field, not part of the name | `checkout-support`                              | `my-agent`, `refund-agent-haiku`  |
| Behavior, `Behavior(name=)`                 | a verb phrase from the team's policy or spec, snake\_case, what a pass looks like                                    | `refunds_when_eligible`, `escalates_over_limit` | `test1`, `quality`                |
| Version, `tracked.run("…")`                 | what the team already calls a release: git tag, PR, date or prompt label                                             | `2026-09-20`, `pr-412`, `policy+dates`          | `v1`, unless the team ships as v1 |
| Harness variant, `Harness(label=)`          | `prompt@model`; the prompt part is the prompt file or the change                                                     | `policy+dates@claude-sonnet-5`                  | `variant-3`                       |
| Test, `test_version=`                       | `"t-" + sha256(asks)[:8]`, the SDK's one fixed convention, because the name must change when the asks do             | `t-5f57ed8d`                                    | `final`, `v2`                     |
| Method, `method=`                           | as the SDK spells it                                                                                                 | `eval`, `GRPO`, `SFT`, `DPO`                    | `training`                        |
| Experiment, `tracked.experiment(question=)` | the question in the team's words, one sentence                                                                       | "Does the dates rule cut wrong refunds?"        | "Experiment 3"                    |

Three checks before the first post:

* A teammate who has never opened the platform would recognise the agent id.
* Every behavior name appears, in some form, in the repo's own docs or tests.
* Versions sort the way the team's releases sort.

Why the test is the exception: a score is comparable only with the setup held constant (rlhfbook.com, "Evaluation"), so the test's name is its content and changes on its own when an ask changes. Everything else is a label a person reads, and the person is on the team.

The `AGENTS.md` block that `whileai init` writes carries this in one line: *you know this repo best; name the agent, behaviors, versions and experiments in its words, and the test by its content.*