Autonomy Governance

1,003 ticks. $0 earned. Nothing ever crashed.

That is the telemetry from a real autonomous agent company with a live crypto treasury, seven specialist roles, and working tools. Every tick "succeeded." Its zero-progress streak reached 964 consecutive iterations before anyone noticed the loop had been optimizing an unreachable metric the entire time.

The agents were not broken. The governance was.

This skill is the forensics and the fix — the control plane that decides what your agents optimize, what they may not touch, and how you know whether any of it actually worked.

What it covers

• The self-reinforcing charter — the 964-tick trap. Why operator edits silently vanish after one tick, and why your agents are obeying a policy you thought you deleted.
• North-star capture — how a perfectly sensible KPI can be structurally unreachable, and the one counter that makes it impossible to miss.
• Measure-only theater — ticks that observe, report, change nothing, and log green. Includes a real tick goal that is 40 words long and accomplishes nothing.
• Decorative escalation — an SLA detector wired to a handler that its own trigger condition disables. It fired true for hundreds of ticks and nothing happened.
• Blast radius — the sandbox pattern that kept the same company from ever touching a live k8s cluster it had full SSH access to.

What you get

• SKILL.md — 5 named failure modes with real code, real telemetry, and the exact fixes
• The scorecard pattern — binary gates ordered along the causal chain to first value, so the first false is always your real blocker
• An 11-point operator checklist to run before leaving any loop unattended
• A 6-step triage for diagnosing a loop that is already stuck
• 8 named anti-patterns with their tells and what each one costs

Install

unzip autonomy-governance.zip -d ~/.claude/skills/

Then reference it whenever you are designing, reviewing, or debugging a long-horizon agent loop.

Who it's for

Anyone running agents unattended past the point where a human would notice they are stuck: agent companies, swarms, heartbeat loops, overnight batch agents, self-directed research runs. Especially if you already own C-suite personas — this is the layer that keeps them from confidently doing the wrong thing for a thousand iterations.

Provenance: every failure mode was observed in production, not derived from theory. The numbers are that system's actual telemetry at the moment of diagnosis.