𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗔𝗻 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗔𝗴𝗲𝗻𝘁 𝗧𝗲𝗮𝗺
I built a small agent team to replicate my engineering workflow. It handles everything from reading a task to pushing a reviewable branch.
The workflow follows these steps:
- Clarifier: Checks if the task is specific enough.
- Planner: Explores the codebase and writes an implementation plan.
- Implementer: Executes the plan and commits code in chunks.
- Reviewer: Finds bugs and style issues.
- Tester: Runs tests and linters to ensure stability.
The core secret is structured communication.
Agents do not chat with each other. They emit structured tokens like PLAN_WRITTEN or REVIEW_RESULT. The orchestrator parses these tokens to route the next step. This makes the system faster, cheaper, and prevents hallucinations.
Each agent has a specific role and model:
• Planner (Opus): Deep reasoning to map out file paths and exact code changes. • Implementer (Sonnet): Follows the plan or fixes specific review blockers. • Reviewer (Sonnet): Compares code against the base branch. It uses an 80% confidence threshold to avoid noise. • Tester (Haiku): Runs Go commands like vet, test, and lint.
Key design choices for reliability:
- One job per agent: This keeps the logic simple.
- Fixed output contracts: This removes the need for extra LLM calls to extract intent.
- Review-retry loop: The system allows up to two attempts to fix blockers.
- Model optimization: I use expensive models for planning and cheaper models for testing to save costs.
You can automate your repetitive loops. You just have to write them down.
Optional learning community: https://t.me/GyaanSetuAi