Multi Agents are FINALLY working.
Everyone thinks multi-agent = throw 10 AIs at a problem and let them cook in parallel.
That doesn't work in pratice.
When multiple agents write code simultaneously, they make conflicting decisions — different styles, different edge case handling. It's chaos.
Cognition and Anthropic independently discovered the same architecture. One agent writes. The others contribute intelligence.
Anthropic calls it orchestrator-worker. Cognition calls it "single-threaded writes with multi-agent intelligence."
Here are the 3 patterns that actually work:
1. The Self-Review Loop Devin reviews its own code with a separate clean-context agent.
Why? Context rot — models make worse decisions at longer context lengths and The fresh reviewer catches what the coder can't see.
Result: 2 bugs found per PR, 58% of which are severe.
2. Smart Friend
Two frontier models (say Opus 4.7 and GPT 5.5) working together. Different models excel at different subtasks — some debug better, some reason visually better. The unlock is routing to each model's strengths.
I personally use Codex to review code made by Claude Code and it catches critical issues constantly.
3. The Research Swarm Anthropic's version: one lead researcher (Opus) spawns parallel search agents (Sonnet) that fan out, gather info, and report back. these workers are read-only
90% improvement over single-agent. 15x more tokens but handles parallel research beautifully.
The pattern everyone's converging on isn't "more agents." It's the right agents doing the right jobs, with one clear decision-maker.
If you're interested in the full research, follow comment AGENT and i'll send you the links :)