Agent Optimizer

Assess and improve agent definitions using a 5-dimension rubric. Works standalone, after creating a new agent (creating-agents), or after agent-intake incorporates an external agent.

When to Use

• Standalone: "Optimize this agent" or "Review agent quality"
• Post-creation: After authoring a new agent with creating-agents
• Post-intake: After agent-intake incorporates an external agent
• Repo audit: "Audit all agents" (batch)

Optimization Workflow

Step 1: Analyze

Single agent:

bash skills/agent-development-team/agent-optimizer/scripts/analyze-agent.sh agents/<agent-name>.md

Batch (all agents, worst first):

bash skills/agent-development-team/agent-optimizer/scripts/audit-agents.sh agents/

Step 2: Apply the Rubric

Use the five dimensions in references/optimization-rubric.md:

Dimension	Threshold	Action if below
Responsibility precision	>70% actionable	Replace preambles with workflow steps; remove duplicated principles
Retrieval efficiency	0 duplicated paragraphs	Point to skills, do not paste skill content into agent body
Collaboration completeness	100%	Add purpose, required, without-collaborator to every collaborates-with
Classification alignment	0 mismatches	Align type with tools and workflow (e.g. strategic ≠ Bash)
Example quality	100% concrete	Add input + expected output per workflow

Step 3: Fix and Re-run

Edit the agent file, then re-run analyze-agent.sh to confirm grade and status (OK / REVIEW / OPTIMIZE).

Optimization Checklist

• Body < 400 lines (longer suggests skill content duplicated)
• Every collaborates-with has purpose, required, without-collaborator
• classification.type matches tools and workflow (no strategic+Bash; quality agents don't produce artifacts)
• At least 3 workflows with concrete input and expected output
• No full paragraphs copied from referenced SKILL.md files
• Actionable content >70% of body (steps, checklists, commands over philosophy)

Cross-References

• Rubric: references/optimization-rubric.md — full scoring and grade bands
• creating-agents — Authoring standards; run validate_agent.py after changes
• refactoring-agents — Overlap and ecosystem fit when optimizing multiple agents
• Command: /agent:optimize [agent-name] or --all for batch