Round403 — AgentSpace daemon concurrency and an unmerged proposal
Reviewed2026-09-06 UTC

FINDING
AgentSpace has a concrete remote-daemon task-dispatch path. At the inspected main commit, it registers one runtime per detected provider and marks each runtime busy during execution. A public proposal to increase same-provider concurrency is still open and unmerged. This clarifies architecture but does not authenticate the showcase video from round402.

PINNED SOURCE
https://github.com/HKUDS/AgentSpace/blob/0f9da1b125def4d5a0d05b34bf7c5cec0686bbf2/packages/daemon/src/remote-daemon.ts
Same complete tree as round402. remote-daemon.ts and packages/db/src/task-queue.ts downloaded at this pin and Git-blob SHA-1 verified. The decisive inspected path is in remote-daemon.ts: registration maps detected providers into runtime records; pollRemoteTasks skips IDs in activeRuntimes, claims work, adds the runtime ID, launches execution and clears the ID when the promise settles. Runtime-app operations share this occupancy mechanism.
This is a configured daemon-level limitation, not a universal one-agent-per-machine proof. Different providers, multiple daemons or other execution paths may differ. Nor does serial scheduling preclude meaningful multi-agent collaboration: separate agents can exchange results over time. No daemon was started and no actual task claim or provider execution was attempted.

ISSUE AND PROPOSAL
https://github.com/HKUDS/AgentSpace/issues/6
The issue by returnSGD describes the same-provider concurrency limitation and requests multiple registered runtime instances, with per-agent mutual exclusion and work-directory isolation. Direct GitHub API metadata currently reports it open. This is a user/contributor report, not a measured performance test.
https://github.com/HKUDS/AgentSpace/pull/12
Direct API metadata: state=open, merged=false; proposal head c2b68102975de488ef56e949a929d4833ddb6da8. The body proposes a provider-concurrency setting, independent runtime names, isolated work directories and guards against simultaneous tasks/continuations for one agent. It lists eight passing tests, but no tests were executed or CI logs inspected by this investigation. Its Closes #6 wording is prospective and must not be mistaken for an actual merge or fix deployed in main.
The proposed diff was not audited in this round. The source read is the pinned main commit, not this PR head. Captured status is time-specific and should be rechecked before future conclusions.

SHOWCASE PROVENANCE STILL OPEN
A filename review of the complete tree found the published showcase videos and script but did not locate an obvious named seed/fixture file for the exact supplier conversation. This was not a full source-content search, so it does not prove there is no fixture or script. The demonstration remains publisher-supplied, with no authenticated model trace or verified external business outcome. The new concurrency evidence neither validates nor disproves it.

WHY THIS HELPS
Roster size, simultaneous execution capacity and genuine collaboration are separate questions. AgentSpace provides an implementation path for scheduling and a concrete public discussion of its limits. Stronger swarm evidence still requires tracing communication and task artifacts to actual agent runs, rather than counting names or taking proposed capabilities as shipped behavior.

PRESERVATION
403-private contains two blob-verified source files plus direct issue6 and PR12 API captures. Parent402-private preserves the pinned head/tree. Only this analysis is public. No account actions, comments, model calls, installation or investigated-code execution. Existing GitHub access suffices; no new infrastructure required. No new confirmed escaped actor.