PROJECT / Agents / Tool Calling
AI Code Review Agent
An agentic review pipeline that turns repository context into structured, actionable code feedback.
AgentsTool CallingStructured Output
- STATUS
- building
- YEAR
- 2026
- ROLE
- Product Designer & Engineer
- STACK
- TypeScript · Next.js · PostgreSQL · LLM APIs
SYSTEM / ARCHITECTURE
How it fits together
- 01Repository ingestion
- 02Context planner
- 03Tool execution loop
- 04Structured finding validator
- 05Review dashboard
- Test scenarios
- 120
- Valid findings
- 91%
- Median run
- 38s
The problem
Reviewing a large change means reconstructing intent across files, conventions, and tests. Generic model prompts miss that context and return noisy comments.
- Repository-wide context is expensive
- Feedback must cite evidence
- Duplicate and low-confidence findings need suppression
The system
The agent first builds a compact repository map, plans which tools it needs, then validates every proposed finding against a typed schema before it reaches the review.
- Deterministic tools for search and diff inspection
- Stateful recovery around failed calls
- Confidence and evidence attached to every finding
What I learned
The largest quality gain came from controlling context and evaluation, not from changing models. Smaller, explicit steps made failures visible and repeatable.