Multi-agent systems split work across an orchestrator and specialized subagents, each with its own context and tools. They can add breadth and parallelism, but every handoff is a trust boundary where context gets lost, errors compound, and one compromised agent can mislead the others.
Read in this order
- ArticleHow we built our multi-agent research system (Anthropic)Start here: a real multi-agent build, with what worked and what did not.
- ArticleDon't Build Multi-Agents (Cognition)The counterargument: why context loss between agents makes many designs fragile.
- ArticleHijacking multi-agent systems in your PajaMAS (Trail of Bits)Shows how one hijacked agent can take over the others in a system.