If AI can generate code in 30 seconds, why does it still take three days to ship a feature? Because of the Pull Request Bottleneck. When human developers are asked to review hundreds of lines of AI-generated syntax every day, review queues stall. Here is how modern teams solve it using Matt Pocock's Dual-Axis Review and Compounding Retros.
Never ask the same agent in the same context to review its own code. Confirmation bias and context saturation ensure it will miss its own hallucinations. You need a fresh, isolated subagent evaluating two distinct axes.
1. Why Self-Review Fails: Context Degradation
A common rookie mistake with AI coding tools is asking the assistant: "Now review the code you just wrote and fix any bugs."
This almost always fails. In modern LLM architectures, an agent that spent 40 tool calls and 50,000 tokens writing a feature has developed strong context confirmation bias. It believes its assumptions are correct. To get an honest, ruthless code review, you must spin up a separate reviewer agent in a completely fresh context window.
2. Matt Pocock's Dual-Axis Review Framework
In Matt Pocock's open-source skill architecture, effective code review does not look at code as a generic blob. It evaluates diffs along two independent, parallel axes:
| Review Axis | The Question It Asks | What It Evaluates |
|---|---|---|
| Axis 1: Standards | "Does this look like our code?" | Repository style, naming conventions, modularity (<300 lines), security rules, error handling, and avoidance of code smells. |
| Axis 2: Spec Coverage | "Did it actually build what we asked for?" | Checking the diff against the original requirement. Did the agent hallucinate extra features? Did it drop edge cases? Are the tests asserting real contracts or just tautologies? |
3. The 3-Tier Layered Quality Model
To eliminate the PR bottleneck without sacrificing quality, Pocock establishes a clear quality hierarchy:
- Tier 1: Cheap CPU Checks: Formatters, linters, TypeScript compiler checks, and unit tests. If these fail, no human or AI reviewer ever sees the code.
- Tier 2: Automated AI Reviewer Subagent: A fresh subagent runs the Dual-Axis Review against
AGENTS.mdand the original prompt. It flags deviations side-by-side. - Tier 3: Human Review for "One-Way Doors" Only: Human architects spend zero time checking syntax or formatting. Human eyes are reserved exclusively for irreversible decisions: database schema migrations, auth security boundaries, and core business models.
4. The Compounding Retro Loop (/retro)
The most powerful concept in Pocock's playbook is the Compounding Retro Loop. In traditional software, when a senior engineer spots an anti-pattern in a PR, they leave a comment: "Don't use raw fetch here; use our apiClient helper." Two weeks later, a different developer makes the exact same mistake.
In the Antigravity workflow, whenever a human catches a mistake during review, you run a retro step:
# The Compounding Rule
"Every human PR review comment must result in an update to your local .gemini/rules/AGENTS.md or RULE[user_global]."If the agent forgot to handle rate limits, you don't just fix the code. You add a single line to AGENTS.md: "Always verify API client calls against the 429 rate limit backoff helper."
From that moment on, every agent invocation automatically obeys the rule. Your engineering standards compound and get smarter with every single commit.
Hostinger
Premium web hosting made easy. Blazing fast speeds, 99.9% uptime, and 24/7 support for your mission-critical applications.
HOST open_in_new
Signal Feedback
No signals yet. Be the first to transmit.