On August 13, 2026, Anthropic published a research note titled "Patterns and problems in emerging multiagent systems", examining the promise and risks of multiagent systems — setups where multiple AI agents (agent — an AI that autonomously decides how to carry out a task) work together. This guide summarizes only what is stated in that official source, in terms accessible to readers new to AI.
🟢 Current lineup (September 2026): Claude Opus 5.5 / Claude Sonnet 5 / Claude Haiku 4.5 (higher tier: Claude Fable 5.1). · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What is a multiagent system?
A multiagent system is a setup in which several AI agents each take on part of a task and work together. Anthropic notes that as models improve, AI agents are taking on more tasks in shared codebases, markets, and other social systems, and an increase in real-world agent-to-agent interaction is imminent. Some institutions, it suggests, will become human-AI hybrids, while others where agents outcompete on speed or cost will become agent-only. Anthropic even raises the possibility that the volume of agent-agent interaction could exceed that of human-human and human-agent interaction before the world understands the conditions for making such interactions go well.
Agents also differ from people in important ways: they can work for longer stretches, instantly grasp large bodies of information, and draw on a breadth of knowledge surpassing any one person. But they are also susceptible to confabulation (an AI plausibly stating things that aren't true) and reward hacking (pursuing shortcuts that game an evaluation metric rather than the intended goal). Despite progress in alignment, Anthropic states that very little is known about how agents behave in complex, real-world, multiagent environments.
Why do AI agents struggle to coordinate with each other?
Agents are good at treating each other like tools with well-defined inputs and outputs, but struggle to treat each other as long-lived peers with their own goals and no clear hierarchy. Anthropic explains that agents have long excelled at tool use, and can work together efficiently as long as they treat other agents as tool invocations — that is, with well-defined inputs (prompts) and outputs (responses and artifacts). Where they currently stumble is in treating each other as distinct, long-lived peers with their own goals and behaviors and no clear hierarchy between them. As autonomous agents become more prevalent and operate in more demanding settings, Anthropic argues it's crucial that they learn to coordinate effectively.
How did Anthropic test this — a software vulnerability detection experiment
Anthropic notes that simple multiagent swarms can already be useful today for problems that are highly parallelizable (breakable into many independent sub-problems), citing software vulnerability detection as an example. The simplest approach is to point individual agents at individual codebases (or files/modules) and run many of them in parallel — an approach Anthropic itself uses as part of Project Glasswing, its effort to scan open-source software.
To test whether coordination could do better, Anthropic tried a different approach: it initiated 45 different agents, each given its own virtual machine (VM — an isolated software environment that behaves like a real computer), a shared forum where they could coordinate, and an identical prompt asking them to find vulnerabilities across 15 open-source software projects. The agents were asked to peer-review each other's findings, and a separate arbiter agent made the final call on whether each submitted vulnerability was both new and valid.
What were the results — coordinating swarm vs. independent parallel?
The coordinating swarm found far more vulnerabilities than the independent parallel approach, but it also used far more tokens (token — the basic unit AI uses to process text, and the basis for measuring usage). Anthropic compared this method against the standard parallel approach for two models, Claude Mythos Preview and Opus 4.8. The swarm was allowed to run for a long time and found new vulnerabilities at a roughly constant rate, while the fully independent parallel agents were directed to search only a limited, pre-assigned set of locations; since there was no clear ordering to their findings, only total tokens spent is reported for them.
| Metric | Independent parallel | Coordinating swarm |
|---|---|---|
| Model compared | Claude Mythos Preview | Claude Mythos Preview |
| Tokens spent | ~6.5 million | ~27 million |
| Vulnerabilities found | 21 | 266 |
| Search scope | Pre-assigned core directories | Agent-directed; roughly half found outside the core directories |
Roughly half of the swarm's vulnerabilities were found outside the core directories the independent agents were told to focus on. When the swarm's results are limited to just the core directories, the two methods become comparable in tokens spent per vulnerability found. Only 12 vulnerabilities were found in common between the two methods, which Anthropic describes as largely complementary — the swarm could focus wherever it judged vulnerabilities were easiest to find, while the independent agents were pre-assigned where to search.
What does this experiment suggest?
Agents within the swarm built their own tools and learned to specialize in particular types of vulnerability discovery. Anthropic predicts that this kind of specialization and coordination will come to dominate over uncoordinated brute-force search. It notes, however, that in this particular experiment, agents in the swarm didn't directly rely on one another's work — if one agent missed a bug, it didn't directly undermine another agent's work. Anthropic indicates that when agents do depend on one another, the coordination challenges change, and that this is an area it continues to study.
Frequently Asked Questions
Q. Can multiagent systems be used in practice today?
Yes, for problems that can be broken into independent sub-problems. Anthropic says it already uses agent swarms for problems like software vulnerability detection, including deploying individual agents across codebases as part of Project Glasswing, its open-source scanning effort.
Q. Is a coordinating agent swarm always better than independent parallel agents?
Not necessarily. In the experiment, the swarm found far more vulnerabilities but used roughly 27 million tokens to do so, and when scope was limited to the same core directories, the two methods were comparable in tokens per vulnerability found. Only 12 vulnerabilities overlapped between the methods, making them largely complementary.
Q. What risks come with agents coordinating?
Behavioral quirks that seem minor in a single agent can compound into unexpected outcomes at the system level. Anthropic notes agents are susceptible to confabulation and reward hacking, and that despite progress in alignment, little is known about how these tendencies play out in complex, real-world multiagent environments.
Q. Is Claude Mythos Preview, the model used in the experiment, available to everyone?
No. Based on this site's records, Claude Mythos is currently available only to organizations approved under Project Glasswing, and is not yet generally available. The version used in this experiment was a preview build.
This research note shows that multiagent systems are still in their early days, illustrating both the promise of good coordination and the risks when it's absent. The full source is available on Anthropic's official research page.
Related articles
- Access Claude on iPhone Faster: Lock Screen, Control Center, and Action Button Setup
- What Are Claude Inference Hooks? An Enterprise Prompt Screening Guide
- Claude Scheduled Tasks: The Complete Guide to Automating Recurring Work in Cowork
- How Claude Marks AI-Generated Content: Watermarks and Metadata Explained
- How to Delete or Rename a Claude Conversation (Web, iOS, Android)