Multi-agent
An architecture where several specialized agents collaborate on the same task — each handles a sub-goal and they coordinate through a shared workspace.
Single-agent systems hit a ceiling around the complexity of tasks they can handle in one trace. Multi-agent splits the work: a planner decides the overall approach, specialist workers handle sub-tasks, and a reviewer checks the final output.
In 2026, multi-agent is mainstream for coding (multiple agents on one PR), research (planner + researcher + writer), and customer support (router + specialists per domain).
The hard part is coordination. Naive multi-agent often over-talks, duplicates work, or loops on disagreements. The teams that get it right impose tight communication budgets and clear hand-off rules.
Frequently asked
Are multi-agent systems always better than a single agent?+
No. For most tasks a well-tuned single agent wins on cost and latency. Multi-agent shines on tasks that genuinely decompose — research with separable sub-questions, codebases with isolated subsystems, support routing across domains.