Patterns and problems in multiagent systems
Models are improving and AI agents are taking on more tasks in shared codebases, markets, and other social systems. As a result, an increase in real-world interactions between agents is imminent. We've already begun studying this, but still have a lot of uncertainty regarding what this looks like at scale. The trajectory is easy to…
Opening of the original on Anthropic Research
Summary
Anthropic's research suggests that coordinating AI agents can significantly outperform independent agents in complex tasks like software vulnerability detection. A swarm of coordinating agents found 266 vulnerabilities, compared to 21 found by independent agents in a similar run. The coordinating swarm also developed specialized tools and focused its attention more effectively, though its findings extended beyond pre-defined search areas. This highlights potential systemic failures and the need for understanding agent interactions at scale.
Why it matters
This research demonstrates a potential leap in AI agent capability when they coordinate, moving beyond simple parallel task execution. This affects developers and organizations relying on AI for complex problem-solving. Unlike competitors who may focus on individual agent power, Anthropic explores emergent behaviors in multiagent systems. Future work should focus on how to reliably steer these emergent behaviors and ensure safety in agent-only systems, especially as agent interactions could soon outpace human ones.
Related: OpenAI: First impressions of GPT-6 Astra from developers · Google: Agent Plugins package your skills, tools, and more · Microsoft: GitHub Copilot app for Beginners: Run several agents at once · Meta: An Organizational Second Brain: Building an AI That Learns From Experts · xAI: Ajeya Cotra – "This might be the clearest warning shot we ever get"
Rated low: routine. Worth knowing, not worth rearranging your day for.
- What Does an Agentic Software Engineering Benchmark Measure? Profiling Task Demands and Agent Behaviour Beyond What Category Labels RevealarXiv cs.AI · Web · Sep 1, 2026
- Sam Altman :‘AGI in 2026’, just as Models Start to [Mis]Train ThemselvesAI Explained · Web · Aug 27, 2026
- Claude Fable 5.1 made me a really nice animated pelicanSimon Willison · Web · Sep 1, 2026
- Just a rumour of a bug is enough to find a security exploit these daysSimon Willison · Web · Aug 28, 2026
- Breaking Claude Code Opus 5 Auto ModeSimon Willison · Web · Aug 27, 2026
- Claude Fable AI Is Much Stranger Than The Headlines SuggestTwo Minute Papers · Web · Sep 3, 2026
- State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490Lex Fridman
- How scientists are using Claude to accelerate research and discoveryAnthropic News
- Anthropic partners with Allen Institute and Howard Hughes Medical Institute to accelerate scientific discoveryAnthropic News
- Claude Science, an AI workbench for scientistsAnthropic News
- FaulT-Bench: Towards Benchmarking Network Troubleshooting LLM Agents under Unreliable User TicketsarXiv cs.AI
- Meta Muse Code & Muse Spark Course – Build AI Agents, APIs, and Full-Stack AppsfreeCodeCamp
Questions people ask
- How did coordinating agents perform in vulnerability detection?
- A coordinating swarm of 45 agents found 266 software vulnerabilities, significantly more than independent agents which found 21. The swarm also developed tools and specialized in discovery.
- What is the main challenge for current AI agents?
- Agents excel at tool use but struggle to treat each other as distinct, long-lived peers with their own goals. Understanding these peer interactions is crucial for complex multiagent systems.
More from Anthropic Research 37 more
Page generated Sep 3, 2026. Summaries are Subvolts' own; the story belongs to Anthropic Research.