A PR opens. The agent swarm wakes up, reviews the code like a senior engineer, posts inline GitHub feedback, suggests before-and-after fixes, learns each developer's habits, escalates critical findings by email to Team Lead, and coaches the team with a weekly learning digest.
Missing idempotency key in charge creation. Retrying after a network timeout could create duplicate charges.
How it works
Five autonomous agents. One transparent reasoning loop. Powered by Google Antigravity.
The Orchestrator autonomously picks which tools to call — triage, habit lookup, deep review, escalation. No hardcoded pipeline. Every decision is logged.
Ranks every file by risk in milliseconds. Lockfiles and generated code are skipped. Only meaningful changes reach the reviewer.
Queries each developer's history of past issues. The review focuses on patterns this specific engineer tends to repeat.
Deep semantic review per chunk. Critical, High, Medium, Low findings with exact line numbers and suggested fixes. Posted inline on GitHub.
Critical findings — credential leaks, SQL injection, auth bypass — automatically alert the team lead by email and surface on the dashboard.
Visualized
CodePulse exposes its agent reasoning in real-time. Watch the Orchestrator decide. See tool calls happen. Audit every escalation. This isn't a black-box reviewer — it's a transparent teammate.
Live agent trace
Every decision logged. Every handoff visible.
Starting agent-driven review loop. 3 chunks available, 1 reviewable file detected.
Ranked files by risk. High-risk: server/src/services/userService.ts (credential + SQL patterns detected).
Developer has 4 prior issues in SQL injection category. Focusing reviewer attention on raw query patterns.
Found 2 CRITICAL, 1 HIGH finding in userService.ts. Line 42: 'SQL injection via string concatenation in raw query'.
Critical finding escalated. Email dispatched to team lead via Resend. Dashboard updated.
Review complete. 3 inline comments posted to PR #14.
Built on Google Antigravity's Interactions API with session chaining and environment-scoped agents. The reasoning loop runs on Antigravity, not in our TypeScript.
The agent chooses which of five registered tools to call, in what order, based on what it learns at each step.
Generate a device token, paste it into the Android app, see every PR analysis and agent trace from your phone.