Everyone’s building multi-agent orchestration frameworks. LangChain, CrewAI, AutoGen, AG2 — the space is crowded and converging. The frameworks are commodity.
But watch someone actually run a multi-agent coding pipeline for a week. The framework isn’t where the value accumulates. The value is in the routing table that develops through contact with reality:
- Gemini’s free tier quota is shared across parallel calls. Launch three simultaneous delegates and you burn through it in one burst.
- Codex scopes writes to its working directory. Point it at a git worktree and it tries to write to the main repo path instead. Silent failure.
- OpenCode’s auth tokens expire without warning. The error message is in Chinese.
- Gemini silently promotes optional dependencies to main dependencies in your pyproject.toml. Every single time.
None of this is in any framework’s documentation. It accumulates through running the pipeline, hitting failures, documenting what actually happened, and encoding the fix. A routing table. A fallback chain. A post-execution checklist.
The 150-line script that tries Gemini, falls back to Codex, falls back to OpenCode — that’s disposable plumbing. The document that says “never launch 3+ delegates to the same provider simultaneously” — that’s the asset.
This is why the orchestration layer is consulting knowledge, not a software product. The code is simple. The knowledge of what breaks and why is hard-won and compounding. Every run generates data. Every failure updates the routing table. The framework stays the same; the institutional knowledge grows.
If you’re building multi-agent systems for an enterprise, the question isn’t which framework. It’s whether you have a systematic way to capture what goes wrong and feed it back into routing decisions. A framework without that feedback loop is a demo. A thin script with a rich failure log is a production system.