openclaw-agent-optimize – OpenClaw Skill

---

name: openclaw-agent-optimize slug: openclaw-agent-optimize version: 1.0.3 description: Optimize an OpenClaw agent setup (model routing, context management, delegation, rules, memory). Use when asked about optimizing agents, improving OpenClaw setup, or agent best practices. triggers:

• optimize agent
• optimizing agent
• improve OpenClaw setup
• agent best practices
• OpenClaw optimization

---

OpenClaw Agent Optimization

Use this skill to tune an OpenClaw workspace for cost-aware routing, parallel-first delegation, and lean context.

Workflow (concise)

1. Audit rules + memory: ensure rules are modular/short; memory is only restart-critical facts.
2. Model routing: confirm tiered routing (lightweight / mid / deep) aligns with live config.
3. Context discipline: apply progressive disclosure; move large static data to references/scripts.
4. Delegation protocol: parallelize independent tasks; use sub-agents for long/isolated work.
5. Heartbeat optimization: treat heartbeat as control-plane; move heavy checks to isolated cron/scripts; offer profiles A/B/C and require user choice if removing checks.
6. Safeguards: add anti-loop + budget guardrails; prefer fallbacks over retries.

References

• references/optimization-playbook.md
• references/model-selection.md
• references/context-management.md
• references/agent-orchestration.md
• references/cron-optimization.md
• references/heartbeat-optimization.md
• references/memory-patterns.md
• references/continuous-learning.md
• references/safeguards.md

openclaw-agent-optimize

OpenClaw agent optimization skill — model routing, context management, delegation, and cron best practices.

An installable OpenClaw skill that packages battle-tested agent-optimization patterns. Drop it into your workspace and get instant guidance on cost-aware model routing, parallel-first delegation, lean context management, and more.

🙏 Credits & Inspiration

This skill is heavily inspired by affaan-m/everything-claude-code — a production-ready collection of Claude Code configs evolved over 10+ months by an Anthropic hackathon winner. The core philosophy of tiered model selection, progressive disclosure, parallel orchestration, and continuous learning was extracted and adapted from that work for the OpenClaw ecosystem.

Thank you @affaan-m 🎉

📦 What's Inside

File	Purpose
SKILL.md	Skill entry-point — triggers + quick-start workflow
references/model-selection.md	Tiered model routing guide
references/context-management.md	Context window discipline + progressive disclosure
references/agent-orchestration.md	Parallel-first delegation, split-role sub-agents
references/cron-optimization.md	Cron job patterns and model assignment
references/memory-patterns.md	Daily + long-term memory file design
references/continuous-learning.md	Hooks → instincts → skills pipeline
references/safeguards.md	Anti-loop + budget guardrails

🚀 Install

Via ClawHub (recommended)

clawhub install openclaw-agent-optimize

Manual

Copy the skill folder into your OpenClaw workspace:

cp -r openclaw-agent-optimize ~/. openclaw/workspace/skills/

📖 How to Use

The skill auto-triggers when you ask about optimizing your agent, improving your OpenClaw setup, or following agent best practices. It walks you through a 6-step audit:

1. Audit rules + memory — modular, short rules; memory is restart-critical only.
2. Model routing — confirm tiered routing matches live config.
3. Context discipline — progressive disclosure; large static data → references/scripts.
4. Delegation — parallelize independent tasks; sub-agents for long/isolated work.
5. Heartbeat batching — checks grouped by frequency tier.
6. Safeguards — anti-loop rules + budget guardrails; prefer fallbacks over retries.

🔑 Key Principles (from everything-claude-code)

• Cheapest capable model first. Escalate only after failure or clear justification.
• Parallelize by default. Sequential execution is a code smell.
• Hooks > skills for observation. Hooks fire 100% deterministically; skills fire ~50–80%.
• Progressive disclosure. Metadata always visible → body on trigger → bundled resources on demand.
• Batch LLM calls. Classify multiple items in one prompt, not one by one.

📄 License

MIT