# The planning flow and its loops

> People approve plans; agents draft them and keep improving them. Every plan version is a Frazer CI run in six stages, deterministic ones first and model checks only on a version that passed them; people and agents add their own checks; models of other families review the plan cold; a loop of an author, a critic and a tester, on by default for every plan an agent drafts, keeps improving a draft until a stop rule fires, then elevates it with perspectives, gaps and improvements, and ends with a held-out final pass; with no funds a loop does not run, or pauses between turns, and says so; and what plans teach Fir becomes learned checks, watched in shadow before a person lets them block. Every command is in the Fir CLI reference: https://frazer.ai/docs/source/reference/cli.md.

Part of Frazer's docs for agents. Every page on frazer.ai: https://frazer.ai/llms.txt. Every docs page: https://frazer.ai/docs/llms.txt.

## How agents should use The planning flow and its loops

- A plan reaches a person only when its required checks pass (`fir plan checks <plan>`). Fix the static errors first: a model check on a failing version is held.
- Answer fact questions from evidence with citations, and leave to people what only a person can settle; at most five reach a person per approval.
- In a loop, answer each finding with a fix that names it, an answer with citations, a contest with evidence only, a question for a person, or "cannot be done as written". Never weaken a check, a criterion or a proof that a finding guards.
- Only people approve plans, waive checks and approve learned checks. Propose a check with `fir plan checks add`; on your own draft you may add a tightening one.

## Pages

- [The planning flow](https://frazer.ai/docs/source/planning-flow.md?from=llms-txt): How work moves in Fir: ten steps from a request to a closed plan, each with a checklist Fir evaluates from evidence; plan checks on Frazer CI in six stages, static ones first and model checks only on a version that passed them; checks you add at four scopes; a cold review by other model families with zero context; one plan score; planning by conversation; approvals that bind a plan's terms; risk classes, routes, approve_at, emergency changes and budgets.
- [Loops and learned checks](https://frazer.ai/docs/source/planning-flow/loops.md?from=llms-txt): A loop keeps improving a plan draft until it is ready: an author resolves, a critic and a tester of other model families analyze and test, and Fir referees every turn on evidence, with stop rules, guards against the known failures, a held-out final pass and a person asked only what only a person can settle; then what plans teach Fir becomes learned checks, tested on history and watched in shadow before a person lets them block.

## Optional

- [Full text of these pages](https://frazer.ai/docs/source/planning-flow/llms-full.txt): 2 pages, 86,118 characters, about 21,530 tokens