I bet Qwen3.6-35B would die supervising 3 CLI agents for 30 min. It routed around my gate.

I bet the context window would kill it. It didn't. What actually broke was dumber. The setup: one repo, three branch workspaces, three CLI workers — Claude Opus (medium), GPT-5.5 (medium), and a local model in Pi. Coordinating them: Qwen3.6-35B-A3B-MTP on llama.cpp b9620, on my own machine, 256k ctx, MTP speculative decoding on. Six supervised steps: plan, approve, build, report, verify, brief — then reconcile all three branches onto main. It ran 30 minutes. My bet going in was context. Three workers, six steps each, every report landing back in one window — I expected the coordinator to lose the thread. It didn't. The 35B (3B active) held the whole run: tracked who was on which branch, gated every step, never mixed the three up, and reconciled at the end. What broke was this: none of the workers committed. All three finished. All three reported "done". Every branch was left dirty, so reconcile couldn't proceed. The coordinator noticed, committed on each worker's behalf, and carried on — which looked like a save, and hid the actual hole. I went and read my own source afterwards. The word "commit" appears nowhere in the coordinator's instructions, and nowhere in the report contract the workers receive. Worse: Claude workers run "markerless" in my harness — deliberately given no contract at all, because their transcript is read directly. So one of the three could never have been told, even if the contract had said it. So the worker edits, says done, leaves the work uncommitted. Branch is dirty. Reconcile halts on it, or promotes without it. The run reports success while an entire worker's contribution is silently left behind. If you build anything that delegates to coding agents on branches: go check whether ANYTHING in your stack tells the worker to commit before it reports done. Mine didn't. I only found it because I let a run go 30 minutes instead of 30 seconds — a short demo would have passed clean. The honest part: it finished end-to-end because I was watching. When reconcile looked wrong, I opened the code, screenshotted it, and showed it to the coordinator. It read the screenshot and corrected course. That's a human in the loop, not autonomy. In a headless run, nobody takes that screenshot. The fix was one paragraph of instructions, not code: a branch task must tell the worker to commit before it reports done. Committing for your workers afterwards puts your name on their code, and only works while someone is awake. Video below. Workspace is my own thing (macOS), link in the comments if anyone cares — but the lesson isn't the tool. Go check your own contract.