Agent Autonomy Score

Context

Coding agents can process feedback, edit code, run tests, and open pull requests. Teams still need a practical way to decide how much autonomy a particular change deserves.

Problem

Treating every change the same slows down presentation work and gives risky changes too much freedom. The gate needed to distinguish a copy edit from a persistence migration without pretending to understand the entire codebase.

Approach

The CLI reads an implementation intent, a unified diff, or both. It calculates visible engineering signals and recommends one of three collaboration modes.

• Unsupervised for low-risk work with normal pull-request review.
• Guided Autonomy when a human should approve the approach before implementation.
• Pair Programming when the human should remain actively involved.

autonomy-score --intent issue.txt --diff change.diff

Autonomy Score: 8/10 (High Risk)
Recommended mode: Pair Programming

Implementation

The deterministic scorer checks intent language, algorithmic risk, directory spread, state and persistence changes, auth, billing, pipelines, and whether risky production changes include tests. The output works in a terminal, a pull-request summary, or a CI threshold.

Tradeoffs and limitations

The scorer favors a policy teams can inspect and tune. It will miss codebase-specific risks and flag some safe changes. It routes work to the right review mode; it does not approve a merge.

What it demonstrates

The project turns a vague governance question into a tested developer tool with explicit stopping points for automation. Calibration against merged, corrected, and rolled-back agent changes is the next step.

Technology

• Python
• Agentic AI
• Risk scoring

Verification links