Skip to content

Agentic Engineering Review — an evidence-first review protocol for AI agents

Agentic Engineering Review

Let your AI agent review a software project with evidence, applicability-aware scoring, explicit trade-offs, and prioritized next actions.

简体中文

Status: v0.4 — public preview. The review protocol and underlying methodology are open for real-world use, feedback, and validation. This is an opinionated engineering system, not an industry standard or certification framework.

Review a Project with One Prompt

If your AI agent can access this repository and the target project, start with:

Review <TARGET_PROJECT> using Agentic Engineering Review at https://github.com/Charlie-Wang-03/agentic-engineering-review. First read and follow AGENT_REVIEW_PROTOCOL.md. Inspect actual project evidence, determine which review dimensions materially apply, give me the structured Diagnostic Score and evidence-backed findings, and do not modify the target unless I explicitly ask.

<TARGET_PROJECT> can be:

  • the agent's current local workspace;
  • a local project directory;
  • a GitHub or other remote repository;
  • a private or internal project the agent can actually inspect;
  • a non-Git project source with sufficient accessible evidence.

The canonical, tool-independent execution contract is AGENT_REVIEW_PROTOCOL.md. Reviews are read-only by default.

What You Get

A normal review produces:

  • an applicability-aware 0–100 Diagnostic Score when evidence coverage is sufficient;
  • Evidence Coverage so missing evidence is visible rather than guessed away;
  • the highest-value engineering gaps worth fixing now;
  • accepted trade-offs that should remain as they are;
  • unresolved questions that need more evidence;
  • deliberately deferred concerns where remediation costs more than it is worth;
  • three to five prioritized next actions.

The score is a compact diagnostic summary, not certification and not a benchmark for comparing unrelated projects.

How It Works

Target Project
    ↓
Your AI Agent
    ↓
Agent Review Protocol
    ↓
Evidence → Applicability → Engineering Judgment → Trade-offs
    ↓
Structured Review

The protocol requires the agent to:

  1. resolve what project state it can actually inspect;
  2. understand the project's purpose, users, constraints, and maturity before judging it;
  3. inspect real evidence rather than trusting branding or README claims alone;
  4. mark each review dimension as Material, Relevant, or N/A before scoring;
  5. use NE — Not Enough Evidence instead of inventing certainty;
  6. distinguish real gaps from justified engineering trade-offs;
  7. remain read-only unless the user separately authorizes implementation.

Built for Agent-Heavy Project Work

Agentic Engineering Review is especially useful for:

  • AI-native builders — independent developers and engineers who rely heavily on coding agents;
  • technical product builders — AI product managers, prototypers, and independent AI developers working directly with repositories and agents;
  • FDE / solution engineers — people operating across customer environments, deployment constraints, providers, permissions, data boundaries, and fast-changing project requirements;
  • maintainers who want stronger engineering judgment without introducing a heavyweight compliance process.

The target does not need to be open source. Review dimensions that do not materially apply are marked N/A rather than treated as failures.

Why It Is Different

Evidence before judgment

The agent should inspect implementation, configuration, tests, CI, dependencies, history, runtime evidence, or other relevant project material before making broad claims.

Applicability before scoring

The system does not force every project through the same checklist. A private internal project may reasonably treat openness as N/A; a static documentation project may have no meaningful local/remote runtime boundary.

Trade-offs, not compliance

A non-default choice is not automatically a defect. Material deviations use:

Default → Conflict → Trade-off → Exception → Evidence → Revisit

A proprietary solver, cloud platform, or provider can be the correct engineering choice when the evidence supports it.

Agent-executable, human-accountable

The protocol is written so capable AI agents can execute it directly, but engineering responsibility remains human. The agent produces structured judgment support; it does not replace accountable project decisions.

Diagnostic Scoring

Each applicable dimension receives an applicability state first:

Applicability Meaning Weight
Material Meaningfully affects project outcomes or risk 2
Relevant Worth reviewing but secondary 1
N/A Not meaningfully applicable excluded

Evidence-backed applicable dimensions are scored from 0 to 4 and assigned High, Medium, or Low confidence. Applicable dimensions without enough evidence are NE instead of receiving a guessed score.

An overall /100 Diagnostic Score is issued only when weighted Evidence Coverage ≥ 70%. See the exact rules in the Agent Review Protocol and the human-readable Project Review.

Powered by an Agentic Engineering Methodology

The review system is backed by a methodology that is readable by both humans and agents.

Its meta-principle is Outcomes over Dogma:

Project outcomes come first. Principles guide engineering judgment; they do not replace it.

The nine current principles are:

  1. Open by Default
  2. Agent-Native, Human-Accountable
  3. Portable over Model-Agnostic
  4. User Sovereignty & Privacy by Default
  5. Local-First When Practical
  6. Justified Dependencies
  7. Progressive Usability
  8. Verifiable by Default
  9. Reversible Change

They intentionally mix durable values, engineering defaults, constraints, decision rules, and preferred means. They are applied selectively according to the target project's real purpose and constraints.

Two Ways to Use It

1. External Review — default

Keep this repository outside the target project and invoke it at meaningful checkpoints such as:

  • serious project start;
  • major architecture or dependency decisions;
  • substantial refactors;
  • preparation for public release;
  • important releases;
  • post-incident or post-failure review.

A normal external review leaves no methodology-specific files in the target project.

2. Selective Adoption — when useful

If a review exposes recurring project-specific needs, translate only those needs into the target project's own engineering artifacts. For example:

  • adapt templates/AGENTS.md.template when coding-agent context is genuinely weak;
  • use a Decision Record for a high-impact choice that deserves durable evidence;
  • add tests, validation, safeguards, or documentation because the target needs them—not to improve a methodology score.

Do not copy the full methodology into every project.

Go Deeper

Resource Purpose
AGENT_REVIEW_PROTOCOL.md Canonical agent-executable review procedure and output contract
Principles Durable engineering defaults and values
Decision Framework Trade-offs, exceptions, evidence, and revisit rules
Practices Faster-moving implementation guidance
Governance How the methodology and protocol evolve
Project Review Human-readable review dimensions and scoring rubric

Limits

Agentic Engineering Review is not:

  • a security certification or substitute for a dedicated security audit;
  • a compliance framework;
  • a guarantee of software quality;
  • a universal maturity benchmark across unrelated projects;
  • a requirement to use one model, agent, platform, or stack;
  • permission for an agent to modify a target project merely because a review was requested.

If an agent cannot access important methodology or target-project evidence, it should report the limitation and degrade the review rather than fabricate completeness.

Repository Status

This repository is also one of the system's own test subjects. It intentionally remains Markdown-first, with a small standard-library validator rather than a heavy documentation stack.

Run:

python3 scripts/validate_repo.py

English is the default operational language. Simplified Chinese is maintained as a first-class reading path for the README and core methodology documents. The canonical Agent Review Protocol remains English-first; review output should normally follow the user's language.

The repository is in public preview. Real-project feedback and cross-agent behavioral validation will guide future versions before any v1.0 stability claim.

Contributing

See CONTRIBUTING.md. Contributions should improve real engineering decisions, review reliability, or adoption rather than add process for its own sake. For security-sensitive reports, follow SECURITY.md instead of opening a public issue.

License

Licensed under the Apache License 2.0.

About

Let your AI agent review any software project with evidence, adaptive scoring, and actionable next steps. 让你的 AI Agent 基于真实证据审查任意软件项目,并给出自适应评分与可执行的下一步建议。

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages