LRLearning RoadmapRead, build, evaluate
Checklists

Agent Rubric

Agent Rubric Score each area from 0 to 4. Area Score Evidence : Tools are narrow and typed Tool inputs are validated Tool authorization is enforced Read and write workflows are sep

rubrics/agent-rubric.md1 min read

Agent Rubric

Score each area from 0 to 4.

AreaScoreEvidence
Tools are narrow and typed
Tool inputs are validated
Tool authorization is enforced
Read and write workflows are separated
Write actions require preview and confirmation
Idempotency is considered
Audit logs exist
Agent loop has limits
Failure cases are tested
Task success is evaluated

Strong Agent Evidence

  • Authorization matrix.
  • Tool call logs.
  • Confirmation screenshots or transcript.
  • Prompt injection tests.
  • Agent evaluation dataset.