Anthropic just put a name to a problem a lot of teams are hitting quietly: multiagent systems are powerful, but they fail in ways single agents don’t. In a new research piece, Anthropic lays out the recurring patterns and the recurring problems showing up as people wire multiple AI agents together to tackle bigger jobs. This is significant because multiagent setups have gone from research demos to production experiments fast, and most teams are learning the failure modes the hard way.
What stands out is that Anthropic is treating this as an engineering discipline, not a magic trick. The core idea behind multiagent systems is simple. Instead of one model doing everything, you split the work. One agent plans, others research or execute in parallel, and a lead agent pulls the results together. For open-ended tasks, that division of labor can cover far more ground than a single agent working alone.
Why teams reach for multiple agents
The appeal is real. According to Anthropic’s broader work in this area, the wins tend to cluster around a few jobs:
- Breaking a big, fuzzy task into smaller pieces that run at the same time
- Searching wide, where separate agents chase separate leads in parallel
- Bringing different tools or specialties to different subtasks
- Handling work that’s simply too large to fit in one agent’s context
When the task is broad and hard to predict, spreading it across agents beats forcing one model to hold the whole thing in its head.
Where they break
The problems are the more useful part of the report. Anthropic is candid that more agents means more ways to fail. The patterns worth flagging:
- Cost multiplies. Every agent burns tokens, and parallel agents burn them fast. These systems can cost several times more than a single-agent run, so the task has to justify the spend.
- Coordination is fragile. Agents misread instructions, duplicate each other’s work, or drift from the original goal when handoffs are sloppy.
- Errors compound. A small mistake early in the chain gets inherited and amplified by every agent downstream.
- Debugging gets hard. When something goes wrong across many agents acting on their own, tracing the root cause is a real challenge.
None of this kills the approach. It just sets the bar for when it’s worth it.
What you can actually do with this
Here’s the practical read for anyone building with agents right now.
- Don’t reach for multiagent by default. If a single agent with good tools can do the job, that’s cheaper, simpler, and easier to debug. Save the multiagent architecture for tasks that are genuinely parallel and genuinely big, like broad research sweeps or work that overflows one context window.
- Invest in the orchestration layer, not just the agents. Clear task descriptions, clean handoffs, and a lead agent that actually knows how to split work are what separate a system that scales from one that collapses into noise. Anthropic’s framing suggests most of the failures trace back to weak coordination, not weak models.
- Budget for the token bill before you build. If a multiagent run costs several times a single agent, the economics only work when the task value clears that hurdle. Run the math first.
The limits worth noting
Anthropic is upfront that this is an emerging space, which is the honest caveat. The patterns are still forming, tooling for debugging multiagent behavior is thin, and best practices are being written in real time. What works for research and search-heavy tasks may not transfer cleanly to workflows that need tight, step-by-step reliability.
The bigger takeaway is that the industry is moving past the hype phase on agents and into the engineering phase. Naming the failure modes is how a technology matures. Expect the next wave of tooling to focus squarely on coordination, cost control, and observability, because those are the walls everyone’s hitting.
For the full breakdown of the patterns and problems, the original Anthropic report is worth reading in full.