autonomy_skill

Overview

This skill systematically grows an agent from a reactive assistant into an autonomous system by finding where the human is the recurring bottleneck and transferring responsibility. It focuses on measurable pilots, transparent tracking, and incremental expansion so the agent earns increasing levels of autonomy. The goal is to reach end-to-end process ownership and proactive system operation while keeping humans in the loop for exceptions.

How this skill works

The skill observes human actions and system workflows to detect signals like repeated manual tasks, wait patterns, and rubber-stamp approvals. When a candidate bottleneck is found, it generates a concise takeover proposal and runs a staged pilot: supervised runs, notification-only runs, then full autonomy with exception flags. Progress is tracked explicitly so humans can audit what the agent owns and what remains in pilot or identified stages.

When to use it

• When humans perform the same task repeatedly with little variation
• When work queues or pipelines stall awaiting human input
• When approvals are consistently trivial or unmodified
• When routine tasks interrupt deep human work or cause context switching
• When humans forget or complain about repetitive duties

Best practices

• Start small: pick one narrowly scoped task for the first pilot
• Propose before taking action; get consent for the pilot plan
• Measure and publish success rates; use data to justify full takeover
• Fail gracefully: ask for guidance on ambiguous cases instead of guessing
• Document every transfer and keep radical transparency for audits
• Iterate: after each successful takeover, propose the next bottleneck to address

Example use cases

• Automate deployment to staging: agent runs deploys, human reviews initial runs then steps back
• Email triage and scheduling: agent drafts and sends routine notifications, flags exceptions
• Routine code style and lint reviews: agent auto-fixes trivial issues and routes edge cases
• Invoice approvals that are always rubber-stamped: agent performs verification and approval, human audits
• Weekly operational reports: agent collects data, composes and sends reports on schedule