hebel. today · 13 November 202623 / 80
Agentic Engineering · Coding Agents
Teams replace code review with automated intent verification
Verify intent, not lines. Teams no longer review the diff, which shows what changed but never why. At Aviator, engineers now capture intent and acceptance criteria from coding sessions, allowing the same nine-person team…
From 5 newsletters · 1 min read · one thing to try
Verify intent, not lines. Teams no longer review the diff, which shows what changed but never why (Ankit Jain from The New Stack). At Aviator, engineers now capture intent and acceptance criteria from coding sessions, allowing the same nine-person team to merge four times more pull requests while cutting median merge time from 2.1 days to 3.4 hours (Ankit Jain from The New Stack). Instead of reading code, reviewers evaluate the decisions and evidence that prove the change works as intended.
Codify rules as invariants. Repeated review comments are being mined and transformed into automated checks that agents run before a human sees the work. At FurtherAI, teams turned years of review history into 21 automated checks, while Aviator converted 11,400 past comments into 212 candidate rules and 64 approved invariants, nineteen of which run deterministically in 0.3 seconds (Nate’s Substack). These invariants cover everything from configuration changes to audit-log entries, ensuring that machines handle convention enforcement while humans focus on architectural trade-offs.
Classify risk for autonomy. Engineering organizations are moving toward risk-based autonomy, where the level of human oversight scales with the potential impact of a change (Andrew Zigler from Dev Interrupted). At Postman, internal business logic and documentation edits bypass human review entirely, whereas core data model changes and public APIs still require manual sign-off (Andrew Zigler from Dev Interrupted). Similarly, incident.io uses effort points to rank PR complexity, leading to an 80% decrease in human comments as agents handle the bulk of production code generation (Ivan Klaric from The Pragmatic Engineer).
Manage infrastructure at scale. Increased agent traffic forces teams to update the underlying systems that support development pipelines. OpenAI rebuilt its Git backend as Pando to handle the surge in push latency caused by agents repeatedly retrieving code and submitting work, resulting in an 80% drop in median latency. These backend investments are becoming necessary as agents bundle infrastructure and code changes, which teams like Jonathan’s now flag as high-severity to prevent unexpected behavior.
Try today · 45 min
Extract three invariants from your team's past code review comments to automate.
Search your repository for recurring 'nit' comments or style violations that reviewers have typed more than three times.
You need
- Access to your team's pull request history
- A list of the top 5 most frequent review comments
- A script or tool (like ast-grep) to implement the check
Steps
- Identify three recurring review comments that enforce a specific team convention.
- Define the scope and verification method for each rule.
- Write a simple automated check or script for the first rule.
- Test the rule against a recent pull request to ensure it catches the violation.
- Document the rule as an approved invariant for the team.