Skip to main content

Goals and rubrics

还没翻译译自英文（docs.langchain.com）本页 30 段都还没译

Goals and rubrics help Deep Agents Code check whether its work satisfies the criteria you care about. Use a goal for objective-driven work where the agent should help define acceptance criteria, and use a rubric when you already know the criteria.

Choose a goal or rubric

Use a goal when you have one measurable objective and want Deep Agents Code to draft acceptance criteria before it starts. A goal has a lifecycle: once accepted it stays active across turns until you pause it, the agent marks it completed or blocked, or you clear it. You can also amend an active goal without restarting the work.

Use a rubric when you already know the criteria you want the agent graded against. A rubric can apply to the next turn only or persist across future turns.

Common patterns:

  • One objective, agent drafts criteria: Use /goal <objective>.
  • Persistent criteria for every turn: Use /rubric set <criteria>.
  • Criteria stored in a file: Use /rubric file <path>.
  • One-turn quality gate: Use /rubric next <criteria>.

Use a goal

Use /goal when you know the outcome you want, but want Deep Agents Code to propose the acceptance criteria before work begins. Goals are useful for open-ended work: the agent turns an objective into a concrete definition of done, then iterates until those criteria are satisfied.

/goal add OAuth refresh handling

Deep Agents Code drafts acceptance criteria for review before starting the task.

How draft criteria are handled depends on your approval mode:

After the criteria are accepted, the goal stays active across turns until it is paused, completed, blocked, or cleared.

This approach lets you work toward a larger objective over multiple turns:

/goal migrate auth callbacks to the new API
start with the OAuth callback
now update the tests
check the docs too

The goal panel above the input shows the current objective and whether it is active, paused, blocked, or completed. Use /goal show to inspect the current goal, and /goal clear to remove it.

Amend, pause, and resume a goal

Steer an ongoing goal without cancelling the current task and replaying work:

  • /goal amend <feedback> proposes coordinated updates to the objective and criteria. The amendment goes through the same inline review (accept, edit, revise, or cancel) before finalizing.
  • /goal pause saves the goal without letting it drive work or grading, so intervening prompts run without it. /goal resume reactivates the saved goal and continues from the existing conversation.
/goal amend remove JSON export, add streaming CSV support, keep the CSV tests
/goal pause
/goal resume

Completion and grading

Each follow-up turn is graded against the goal's acceptance criteria until the work is done.

  • When a goal's completion is approved, Deep Agents Code clears the goal

Use a rubric

Use /rubric when you already know the acceptance criteria and want them to act as a quality gate for the agent's work.

/rubric set tests pass; no unrelated files changed; help text is updated
/rubric next only change the auth callback; do not refactor unrelated code
/rubric file acceptance.md

A sticky rubric applies to future turns until cleared. A next-turn rubric applies only to the next submitted task.

See also

AI 翻译查看英文原文