The Multi-Agent Problem Nobody Talks About
Every AI startup pitches multi-agent orchestration as the future. AutoGPT. CrewAI. LangGraph. OpenClaw. The pitch deck always shows three or four specialized agents working in concert — a coder, a researcher, a customer support bot. The pitch deck never shows what happens when those agents share a single global config and step on each other's settings every time someone switches context.
That's what OpenClaw shipped. Multi-agent support launched over a year ago to genuine excitement. What the launch post didn't mention: every agent inherited the same thinking mode, the same reasoning visibility, the same fast-mode toggle. If you wanted Agent A to think deeply and Agent B to respond fast, you toggled manually. If a stored model override conflicted with a different agent's allowlist, the system silently did the wrong thing. No warning. No fallback.
Koc's PR #51974 adds what should have been there from day one: per-agent defaults for thinkingDefault, reasoningDefault, and fastModeDefault. The resolution chain is sensible: per-agent overrides sit between session-level and global config. Stale model overrides auto-clear with a system notification. It's exactly the boring, correct implementation the feature needed at launch.
Discord's Double Failure
The two Discord fixes are related but discovered independently. dutifulbob found that abort signals from timed-out Discord workers never reached the ACP dispatch layer — agent sessions kept running in the background, burning compute, after the Discord side had given up. The fix was two lines threading the signal through. The plumbing already existed. It just wasn't wired.
Takhoffman found something worse. The dedup guard marked messages as “seen” before preflight validation ran. If a transient failure occurred, the retry was suppressed for the entire TTL window — five minutes of “I already handled that” for a message that was never actually handled. At-most-once delivery by accident. The kind of bug that makes operators think Discord is unreliable when it's the middleware lying.
“The dedup guard achieved exactly one thing: it prevented the system from recovering from the failures it was supposed to guard against.”
The Session Alias That Broke /new
BryanTegomoh's fix is the kind of bug that makes you wince. The /new startup flow — the first thing every new user does — called session_status(sessionKey="current"). The system didn't recognize “current” as a valid key. First impression: an error. The fix took seven commits because the initial approach was a one-off patch; the reviewers (correctly) insisted on normalizing “current” at the shared resolution layer.
This PR was open since issue #39570. How long ago was that filed? Long enough that a seven-commit fix went through 24 test cases and 37 CI checks before merging. Long enough that new users had been seeing that error for months.
The Industry Lesson
Here's what the broader AI agent industry should take from this: multi-agent is a feature that sells demos and a capability that demands infrastructure. Per-agent config isolation, session identity management, abort signal propagation, idempotent message delivery — these are table stakes for any system running multiple concurrent agents. OpenClaw shipped the demo first and the infrastructure fourteen months later. They're not the only ones.
Four PRs. Four contributors. Four gaps in the foundation, all merged the same day. The multi-agent story is stronger now. But the next time someone shows you a multi-agent demo with specialized bots working in harmony, ask them one question: can those agents have their own settings? If the answer involves the word “global,” walk away.