5 research-backed coding agent orchestration patterns — and when multi-agent hurts by 39-70%
BorderlineImmigrant · reddit · 2026-11-22
A practical guide by YK Sugi (Google Cloud) compiles empirical research on multi-agent coding orchestration: up to +80.8% gains on cleanly independent subtasks, but 39-70% performance drops on tightly coupled planning tasks.
The five patterns:
- Builder + Verifier: a read-only agent in a fresh thread reviews diffs, removing same-context self-review bias. Good default for non-trivial changes.
- Split into Lanes: parallelize only when file ownership divides cleanly; overlapping edits cut success rates by 30%.
- Best of N: run the same hard task N times, let automated tests pick the winner — pass rates jumped from 15.9% to 56% on real GitHub issues, but only with objective automated checks.
- Council of Models: for open-ended work without deterministic tests, gather independent proposals from different models and have blind judges score them.
- Team of Reviewers: parallel read-only reviewers each own a single lens (UX, simplicity, edge cases, security), catching 35% more defects than general checklists.
Includes a routing heuristic: single thread for simple tasks, Builder+Verifier for medium, lanes for independent modules, Best of N for hard tasks with tests, Council of Models for open-ended, Team of Reviewers for large risky PRs.
More from coding & agent
- Harrison Chase endorses 'company ontology' as the right frame for agent-ready company knowledge — hwchase17 · 2026-11-22
- Building a full game on Gemini 4 Neon entirely from a phone: 'shocking and delightfully fast' — aziz4ai · 2026-11-22
- Redditor makes an animated show starring Claude, coded by 20k+ lines of Claude-written Python — quaisar3 · 2026-11-22
- YC's Garry Tan Touts GBrain: AI Workspace That Wrote 850+ Pages of Team Memory in 3 Weeks — garrytan · 2026-11-22
- Overview of 16 LLM eval frameworks: what they score and how they gate — blaizedsouza · 2026-11-22
- Apify CEO: Agents Now Pick Tools for Users, and Trust Becomes the Bottleneck — aisatsana__ · 2026-11-22