Swarms That Refuse Up Front
A headless Swarm aimed at an agent with no locked-down sandbox mode used to be accepted and then die silently — every worker gone within a second, no output, no explanation. That request is now refused the moment you make it, with a message naming the agents that can actually run sandboxed Swarm workers (Codex, Claude and Cursor) and pointing everyone else at terminal Swarms, direct delegation or an Agent Team. Delegating agents get the same list in their orchestration instructions, so they stop composing swarms that cannot start. And a run that dies before its agent ever launches now records why — agent not installed, sandbox not supported, launch refused — in the run record you can read in the Orchestration dashboard and ship with a bug report.
Download v1.57.1
The download page always has the newest build; installers for this version are on GitHub.
What changed
Improvements
2- A Swarm that cannot run is refused up front, with the reason — a headless Swarm targeting an agent with no locked-down sandbox mode (Grok, Antigravity and most others) used to be accepted and then every worker died within a second, with no output and no explanation anywhere. The request is now rejected immediately with a message naming the agents that can run sandboxed Swarm workers — Codex, Claude and Cursor — and pointing to terminal Swarms, direct delegation or an Agent Team for everyone else. This was reported from a Windows machine but affected every platform the same way: sandboxing is about the target agent's abilities, not your operating system. Delegating agents are told the supported Swarm targets in their orchestration instructions too, so they stop composing swarms that cannot start — the instructions refresh on the next app start.
- Failed orchestration runs now say why they failed — a run that dies before its agent ever starts (agent not installed, sandbox not supported, launch refused) records the reason in its run record, visible in the Orchestration dashboard and included in the run files you'd share when reporting a problem. Until now those runs exported as a bare "error" with nothing to diagnose, which is exactly what made this bug hard to report.