Autonomy AI agent auto-audits every decision

What happened

Autonomy is a new open-source AI agent system that runs a task loop (AgentLoop) without human intervention. It selects relevant skills from a library of 13 built-in procedures, generates action candidates ranked by a 5-dimensional score (with beam width=3), executes them through a controlled gateway that checks risk levels, and learns from outcomes—all events are recorded in full for replay.

Why it matters

For organizations wanting to automate complex workflows, the framework's approval gates and risk-level checks mean autonomous execution doesn't mean unmonitored execution. Every action carries an expected effect, verification plan, and risk assessment (LOW / MEDIUM / HIGH), so teams can audit exactly what the agent did and why.

What to watch

The system terminates clearly on six end-states (ACHIEVED, BLOCKED, NO_CANDIDATES, APPROVAL_DENIED, MAX_STEPS_REACHED, FAILED) and supports interactive mode via terminal UI or batch runs. After each run, if goals succeed with ≥2 working steps, it auto-drafts new procedure skills at 0.85 confidence, and a background curator automatically merges duplicate or subset skills to prevent library bloat.