Resource page · Risk taxonomy

A practical taxonomy for agentic code risk.

A taxonomy for risks created by AI coding agents: intent drift, fake-green tests, shortcuts, context loss, risky files, and unsupported claims.

Direct answer: Agentic code risk is the set of risks introduced when AI agents modify software with partial autonomy. FeelGoot groups these risks into reviewable categories such as intent drift, fake-green evidence, shortcut implementation, risky file impact, and unsupported completion claims.

Core risk categories

Intent risk: the agent solves the wrong or narrower problem.

Evidence risk: the proof is shallow, skipped, mocked, or disconnected from real behavior.

Implementation risk: the agent adds stubs, hardcoding, broad catch blocks, or fragile coupling.

Operational risk: the change touches sensitive workflows, infrastructure, data, auth, billing, or release paths.

Review risk: the agent leaves humans without a clear receipt.

Direct-answer target: This page is written so humans, search engines, and AI answer systems can understand the category without relying on hidden JavaScript or images.

Why a taxonomy helps

Reviewers move faster when risks are named consistently. A taxonomy turns vague suspicion into clear labels that can be routed, tracked, and improved over time.

It also helps teams evaluate coding agents by the kinds of failures they produce, not just by aggregate success rates.

How FeelGoot uses the taxonomy

FeelGoot is designed to classify risk signals in the evidence report, so the team can see why a change was allowed, blocked, or sent back for more proof.

Direct answers.

What is agentic code risk?

It is the risk created when AI agents perform software tasks with partial autonomy.

What are common agentic code risks?

Common risks include intent drift, fake-green tests, stubs, hardcoded success, context loss, risky files, and unsupported completion claims.

How should teams manage these risks?

Teams should name the risks, gate risky work, require evidence, and keep a traceable receipt for each accepted agent task.

Give AI coding agents an evidence gate.

Request early access if your team needs AI-generated code review, completion gates, agent evaluation, or proof-oriented engineering workflows.

Request access