Why 'don't build multi-agent systems' keeps winning the argument on Hacker News

Cognition — the team behind Devin — published a blog post in June 2025 arguing you shouldn’t build multi-agent systems, and it’s still the post practitioners reach for a year later when a multi-agent project stalls. Author Walden Yan’s core claim about parallel subagent architectures: “This is a tempting architecture, especially if you work in a domain of tasks with several parallel components to it. However, it is very fragile.” The failure mode he describes is specific — subagents that each get a fragment of context, make independent decisions from it, and produce outputs that don’t reconcile because no single agent held the full picture.

A more recent Ask HN thread from March 2026 — “How are you using multi-agent AI systems in your daily workflow?” — reads like a field confirmation of that thesis. User jovanaccount names “the biggest underappreciated problem”: state coordination, where two agents silently overwrite shared state and the output still looks reasonable, masking the error until it hits production. User mrothroc makes a related point about where the leverage actually is: “The arrangement of the checks between agents matters more than which model you pick.” User dhruvkar reports straightforward performance degradation running subagents compared to a single agent, and is still looking for a workable shared-memory pattern across instances. User formreply’s fix for shared-thread chaos — agents that “race to add the last word” and converge on “verbose non-decisions” — is blunt: “role clarity + veto rights,” with inbound events treated as hard task boundaries so the system stays auditable after the fact.

None of this is anti-agent sentiment — the same thread has people running multi-agent setups in production, with SQLite-backed job queues and human approval gates layered in specifically to contain the failure modes above. The pattern across both sources is that the frameworks aren’t the bottleneck; the orchestration discipline around shared state and decision ownership is.

What this means if you’re evaluating a framework

If you’re choosing between LangChain, LangGraph, CrewAI, LlamaIndex, or AG2 (formerly AutoGen) for an orchestration layer, the HN thread and the Cognition post both point at the same due-diligence question: how does this framework represent and share state across agents, and what happens when two agents act on stale or conflicting context at the same time? That’s a more useful evaluation criterion than benchmark scores or launch-post claims, and it’s exactly the kind of thing worth checking against a live comparison rather than vendor marketing — see our LlamaIndex vs. LangGraph comparison for how two of the category’s frameworks actually differ on this. The full landscape of agent orchestration vendors is tracked on Crail’s agent frameworks category page.