Most multi-agent setups fail because they have too much process. Ours built 11 libraries in 6 days with less.
retained text
Hot take from a few weeks of running agent teams around the clock: the main failure mode isn't bad models. It's process. Agents love inventing rules, gates and checklists, then dutifully obeying them. Ask for rigor and you get a full end-to-end test run for a one-line change. What we run instead, which has now built 11 from-scratch C# libraries (about 1.14M lines, 7 on NuGet) since last Wednesday: **A design interview the agents answer themselves.** A planner asks the design questions and a project manager agent answers them for me. Every decision records who made it. On the first library I sent 112 messages. By the fourth I sent 11, mostly about names, money and credentials. **One orchestrator per library, with units.** Each unit gets a git worktree, an engineer agent, a memory slice and a written brief. The orchestrator verifies and merges. Engineers can't start agents or message upward, so every brief carries a standing go-ahead. **Handoffs instead of giant contexts.** At 400k tokens an agent writes a handoff note and stops, and a successor resumes from it. Nobody keeps a "long-running agent" alive. **Outside judges, not agent-written tests.** For example: test262 for the JS engine (99.55%), html5lib-tests for the HTML parser (100%), Microsoft's DXC for the shader compiler (330/330), and headless Chrome for the renderer, which is faster on every document we benchmark. **Libraries feed each other.** The HTML renderer depends on our HTML, font, image and PDF libraries. It files requests with their project managers, and fixes ship as releases within hours. The rule we keep relearning: give a flexible agent intent and hard constraints, not keystrokes. (Write-up with the numbers in the comments.)
document fetched 2026-10-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xgv8uj/most_multiagent_setups_fail_because_they_have_too