Phase 3 — Grade → Iterate (the bounded loop)
Phase 3 of building a Claude Managed Agent: runs a bounded grade→iterate loop. It defines a rubric-graded outcome, reads grader verdicts to decide the next move (sharpen, re-run, or promote), and once a version passes, runs held-back eval cases in parallel. Produces payloads and run plans for the CMA loop.
✨ What it does
- Builds a user.define_outcome payload with a required rubric and max_iterations clamped to 1..20.
- Reads grader verdicts and recommends next move: SHIP, SHARPEN, ESCALATE, or RESUME.
- Enforces loop invariant: each iteration must move at least one rubric line fail→pass.
- Generates held-back eval cases and a parallel run plan capped at 25 threads.
- Provides a forcing-question library to guide rubric definition and iteration strategy.
🎯 When to use it
- When the user says 'grade my agent', 'make it pass the rubric', 'iterate until it's good', or 'is it good enough'.
- When the orchestrator routes to phase=grade-iterate.
- When you need to run a bounded improvement loop with a rubric and grader feedback.
- When you need to validate generalization with held-back eval cases after a version passes.
🚀 How to use
Trigger by asking to grade or iterate your agent, or when the orchestrator routes to phase=grade-iterate. Provide a build sheet (JSON) and optionally specify max iterations. Run the three scripts in order: outcome_builder.py to define the outcome, verdict_reader.py after each grader result, and eval_scaffold.py once a version passes. Example prompts:
grade my agent and make it pass the rubric
iterate until it's good, but cap at 5 iterations
📄 Output: Deliverables include an outcome payload JSON, a verdict reading with next-move recommendation, and an eval scaffold JSON with held-back cases and a parallel run plan.
📦 Add this skill to Claude Code
# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills
# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/agent-launcher/skills/grade-iterate .claude/skills/grade-iterateSkill source: agent-launcher/skills/grade-iterate/SKILL.md
⚠️ Good to know
Requires a rubric in the build sheet; max_iterations is always bounded (1..20) and never unbounded.
❓ FAQ
What happens if max_iterations is reached?
The verdict reader recommends ESCALATE when max_iterations_reached or failed, meaning you should escalate now rather than continue iterating.
Can I run eval on cases the agent already saw?
No, held-back cases must never be seen during iteration; otherwise you can't grade generalization.
What is the loop invariant?
Each iteration must move at least one rubric line from fail to pass, or the run halts at the cap and escalates.
🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.