Weeks into building multi-agent review orchestration — models checking models’ work, with receipts — I kept hitting failures that had nothing to do with model quality. Completed turns arriving empty. Dispatched tasks vanishing. Files verified on disk that the app insists don’t exist. Two vendors’ desktop apps. Same shapes, over and over.
If you’re building (or dreaming of building) real multi-agent systems on Codex or Claude Desktop, consider this a warning — and if you’ve already hit it, I’d like to hear from you.
What it looks like
Three failure modes, all observed in the last few days on current builds:
1. Empty success. The agent works for ~80 seconds. The turn reports completed. The reply payload is null and the authoritative re-read shows nothing. Four occurrences, two different frontier models — so this is the host dropping the result, not the model failing to produce one. Each one tempts an expensive rerun. Our rule now: never rerun before re-reading.
2. Vanishing tasks. Dispatch returns a client ID. The task never registers — repeated authoritative lookups find nothing, and no duplicate either. Silent loss behind a successful call.
3. Files that exist but won’t open. Markdown verified byte-for-byte on disk, unopenable from the desktop app. The kicker: I caught the sibling of this bug in the other vendor’s app, on a Windows file with zero WSL involvement — the agent wrote it, the chat shows it, the preview says “couldn’t find this file.”
The twist: the bridge is innocent
My first theory was the Windows↔WSL boundary. The data killed it. File bytes verify cleanly in both directions (full reads, matching hashes). One empty completion reproduced with no cross-boundary provisioning in play at all. The failures sit above the byte layer: message hydration, registration lookup, path resolution. Same shape in three places is a systems bug class, not bad luck. I’ve started calling it “empty success” — success status, no payload — and it is poison for orchestration, because every downstream step trusts the status.
Not just me
This is a known cluster with months of history. Our repro data (current builds, cross-model occurrences) is now on the record:
Codex Desktop can’t preview WSL files — our 26.901 repro:
https://github.com/openai/codex/issues/33773#issuecomment-5562836440
Codex Desktop writes empty assistant messages on continued threads — our 4 occurrences:
https://github.com/openai/codex/issues/28751#issuecomment-5562836571
Task registration loss (new issue):
https://github.com/openai/codex/issues/43301
Claude Desktop preview/attach scoping failure (new issue):
https://github.com/anthropics/claude-code/issues/92564
Worth saying plainly: reporters have been asking for dedicated ownership of this boundary — real end-to-end regression tests, release blockers for break-the-workflow bugs — since spring. Fixes land, then regress. That is a prioritization problem, not a talent problem.
The structural frustration
No model benchmark catches “the turn completed and the payload never arrived.” The labs’ eval culture measures single-turn quality; multi-agent systems live or die on transport contracts, idempotency, failure semantics — boring infrastructure that doesn’t trend. You can ship the smartest model on earth and still block AI systems work if the desktop plumbing silently drops results. The gap between model progress and systems reliability is now the binding constraint on my roadmap, and I doubt I’m alone.
What we’re doing
Staying on the Windows installs — migration prices out worse than mitigation. The mitigations, for fellow travelers: persist verdicts to files (worked when the message channel didn’t), run only lanes that produce receipts, hand the apps UNC paths never raw Linux paths, pin and re-verify after every auto-update, and escalate with repro data instead of complaints. Full reasoning is in the linked filings.
Chime in
Hit empty completions, vanishing tasks, or unopenable files in Codex or Claude Desktop? Reply with app, build, and symptom — Mac reports especially welcome, because right now I can’t tell whether this bug class is Windows-only. And if you’re planning orchestration on desktop task dispatch: budget for this up front, demand receipts from every transport, and never trust “completed.”
Hey Codex!
Guys, you do great work. But this is a major bug blocking important development. It’s real, and it’s a gigantic pain in the butt!!
@OpenAIDevs
@OpenAI
@thsottiaux
@ajambrosino
@romainhuet
@embirico
@dkundel
@Dimillian