How do I install the Control Loop skill?

Run npx skillmds@latest add zedarvates/control-loop in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Control Loop safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Control Loop?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Control Loop free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Control Loop?

zedarvates (@zedarvates) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Control Loop by zedarvates — Agent Skill

Name: Control Loop
Author: zedarvates

Control Loop

Close the system into a self-improving loop — measure routing outcomes (local %, token savings, escalation/success rates) and adapt the effort→tier thresholds the auto-router reads, so it gets better at keeping work local without failing. Use to see routing stats, tune the router from real outcomes, or reset thresholds. Also use when the user wants the system to adapt/learn its routing automatically.

zedarvates 9e945ca 5 files · 12.2 KB Updated Sep 17, 2026

File contents

control_loop — the router that learns

The final systemic loop: every routing decision + outcome is recorded, and the effort→tier thresholds (read live by [[auto_router]]) are nudged from the data — keep more work local when local is reliable, escalate sooner when it isn't.

python -m skills.control_loop.cli analyze            # local %, savings, escalation/success
python -m skills.control_loop.cli adapt              # proposed threshold change + why
python -m skills.control_loop.cli adapt --apply      # write it; the router uses it next time
python -m skills.control_loop.cli reset              # back to defaults

How the loop closes

Measure — auto_router.run records each call to ~/.botte/control-ledger.jsonl (effort, tier, local/cloud, tokens saved, escalated, success).
Analyze — aggregate: local %, escalation rate, success rate, tokens saved.
Adapt — conservative rule, needs ≥10 samples, small steps, clamped:
- local reliable (success ≥85%, escalation <15%) → raise the LOCAL boundary (keep more work local → more savings).
- local insufficient (escalation >30%) → lower it (escalate borderline sooner).
Apply — write ~/.botte/routing-thresholds.json; auto_router.effort reads it live, so the next decisions reflect what actually worked.

This is the Karpathy/second-brain pattern applied to routing: the system gets smarter from its own outcomes. Exposed via [[llm_mcp]] as routing_stats. Related: [[auto_router]], [[conductor]], [[metrics]].

---name: control-loopdescription: Close the system into a self-improving loop — measure routing outcomes (local %, token savings, escalation/success rates) and adapt the effort→tier thresholds the auto-router reads, so it gets better at keeping work local without failing. Use to see routing stats, tune the router from real outcomes, or reset thresholds. Also use when the user wants the system to adapt/learn its routing automatically.---# control_loop — the router that learnsThe final systemic loop: every routing decision + outcome is recorded, and theeffort→tier thresholds (read live by [[auto_router]]) are nudged from the data —keep more work local when local is reliable, escalate sooner when it isn't.```bashpython -m skills.control_loop.cli analyze # local %, savings, escalation/successpython -m skills.control_loop.cli adapt # proposed threshold change + whypython -m skills.control_loop.cli adapt --apply # write it; the router uses it next timepython -m skills.control_loop.cli reset # back to defaults```## How the loop closes1. **Measure** — `auto_router.run` records each call to `~/.botte/control-ledger.jsonl` (effort, tier, local/cloud, tokens saved, escalated, success).2. **Analyze** — aggregate: local %, escalation rate, success rate, tokens saved.3. **Adapt** — conservative rule, needs ≥10 samples, small steps, clamped: - local reliable (success ≥85%, escalation <15%) → **raise** the LOCAL boundary (keep more work local → more savings). - local insufficient (escalation >30%) → **lower** it (escalate borderline sooner).4. **Apply** — write `~/.botte/routing-thresholds.json`; `auto_router.effort` reads it live, so the next decisions reflect what actually worked.This is the Karpathy/second-brain pattern applied to routing: the system getssmarter from its own outcomes. Exposed via [[llm_mcp]] as `routing_stats`.Related: [[auto_router]], [[conductor]], [[metrics]].

Frequently asked questions