← All featuresAI Agent Work Validation

Don't take the agent's word for it. Check the work.

Validate AI coding agent work in Flame IDE. Run ESLint, tests, builds, and custom checks automatically after the agent finishes, with optional AI fixes before you commit.

The validation report in AI Chat with sample project data. Failed checks keep their output, and Fix with AI sends it back to the agent. Select the image to enlarge.
WHY IT HELPS

Let your own checks decide whether the work is done.

An AI coding agent can say that everything passes without proving it. Flame runs your configured checks itself after the agent finishes, reports each one, and holds the commit prompt until they pass or you choose to ignore them. Routine verification stays with ordinary commands, so you spend model time on the work that needs judgment.

Run the same checks every time

Configure validation rules once. After an agent turn, Flame runs shell commands, npm scripts, or an independent AI review, such as ESLint, unit tests, and builds, and shows a status for each rule instead of relying on the agent's summary.

Check what the agent changed

Pass only the files the agent edited to tools that accept file paths, and scope each rule with include and exclude patterns. Unrelated changes already in your working tree are left out of the file list.

Hold the commit until the work is checked

The chat's commit prompt waits for validation to pass or for you to ignore it explicitly. Ignoring a failure never turns it into a pass, and the report stays attached to the turn for later review.

Turn a failure into a repair

When a check fails, send the logs back to the agent with Fix with AI, or enable an optional automatic fix after a delay you choose. Flame reruns the relevant validations after each repair.

AI AGENT WORK VALIDATION IN PRACTICE

Run the checks you configured, every time.

Validation is independent of the model that wrote the code. When a turn that changed files ends, Flame runs the rules that apply to the project and reports the outcome rule by rule.

Rules at the scope you choose

Set rules globally, for a project group, or for a single project. Rules merge by ID, so a project can override a group rule or disable one it inherits. Each rule has a working directory, a timeout, and include and exclude patterns.

Only the files the agent changed

Pass the relevant changed files to a tool such as ESLint with the files placeholder, so a lint check looks at the agent's edits instead of the whole repository. A file-aware check with no matching files is skipped, not passed.

Reports that stay honest

Skipped checks stay visible as skipped. Interrupted runs settle instead of spinning forever, and a report goes stale if the checkout changes afterward. Flame rechecks a passed report before handing off to the commit prompt.

AI AGENT WORK VALIDATION IN PRACTICE

A failed check starts the next step, not a dead end.

Failures arrive with diagnostics: exit code, elapsed time, and the output of each check. You choose how much of the repair loop to automate.

Fix with AI, with the logs attached

Start a repair in the same chat and checkout. The failure logs go to the agent as diagnostic data, and validation runs again when the repair finishes.

Optional automatic fixing

Automatic fixing is off by default. Turn it on at any scope and choose a delay from ten seconds to 24 hours. A countdown shows in the failure card, you can cancel it, and there is at most one automatic repair per ordinary user turn.

Run commands on validation events

Agent hooks can run your own commands when validation starts, passes, or fails. They receive the report as JSON, run in the background, and never change the validation result.

WATCH IT WORK

See the workflow in a short demo.

Official Flame demo · 1:37

Make AI Agents Prove Their Work: Lint, Tests & Typecheck After Every Run

Watch Flame run the configured lint, test, and typecheck rules after an AI coding agent finishes, show each check's log, send failures back with Fix with AI, and present the validation report before you commit.

HOW TO USE IT

From rule to reviewed commit.

  1. Add validation rules

    Open Settings, then AI, then Agent work validation, or use the project or group settings. Add a rule for each final check, such as lint, tests, or a build. Use the files placeholder as a separate argument when the tool accepts file paths, and save the validation editor separately.

    npx eslint -- {{files}}
  2. Turn on Validate agent work

    Keep Validate agent work on in the chat's More menu. It defaults on for new project chats when effective rules are enabled, and the choice is remembered per project. Plan, discussion, and simple-chat turns do not run validation.

  3. Let the agent work and read the report

    After the agent turn, Flame runs the applicable rules in order and continues after a failure so the report is complete. Expand a rule to read its status, elapsed time, exit code, and output.

    Sample report. The ESLint and unit test checks failed, so the commit prompt stays on hold.
  4. Fix, rerun, and commit

    If a check fails, choose Fix with AI, Run validations again, or Ignore validation. Once validation passes or you ignore it, the commit prompt appears with the report attached. Direct Git operations and the Changes panel stay available the whole time.

REQUIREMENTS & LIMITS

Know what the tool depends on.

  • A passing report covers only the checks you configured. Weak or missing rules produce weak validation, so include the lint, test, and build commands that matter to the project.
  • Validation is a chat workflow gate. It does not block direct Git operations or the Changes panel, and plan, discussion, and simple-chat turns do not run it.
  • Projects without enabled rules are listed as skipped, and in a chat with several projects each checkout resolves its own rules. Flame does not validate a skipped project.
  • Shell and npm rules run real commands in your checkout with its environment. Use finite commands: watch-mode commands eventually time out, and each rule has a timeout between one second and one hour.
  • An independent AI review is a second opinion from the chat's provider and model. Missing, oversized, or incomplete evidence fails the review instead of passing it, and it uses model tokens, unlike a plain shell or npm check.
FAQ

AI Agent Work Validation: common questions