Chain of Agentsdefinition and how it works in 2026
- Chain of Agents
- An architecture where agents run in sequence, each refining or extending the previous agent's output. Used for long documents or multi-stage workflows.
Chain of Agents passes work through a sequence of specialized agents — one might extract claims, the next verify them, the third draft a summary, the fourth polish the prose. Each agent has a focused scope; the pipeline gets the depth of multi-step reasoning without overloading any single context window.
The technique solves the long-context problem: instead of stuffing a 200-page document into one agent's context, you let a chain process it in chunks, with each agent reading the previous agent's output rather than the raw source. Google's 2024 paper showed chain-of-agents matching long-context models on summarization at lower cost.
In 2026, chain-of-agents is a common pattern inside agentic-RAG pipelines, deep-research workflows (Perplexity, OpenAI Deep Research, Anthropic Research), and document-processing systems. Different from [mixture-of-agents](/glossary/mixture-of-agents) — sequential refinement vs parallel synthesis.
Frequently asked
When should I use chain-of-agents vs a single agent with long context?+
Chain wins when each step has different requirements (extract vs verify vs summarize) or when the source is too long for a single context. Single agent wins when the work is uniform and fits in one context.
Do all agents in the chain need to be the same model?+
No, and they often shouldn't be. Cheap fast models do extraction and routing; expensive frontier models handle the synthesis steps. This hybrid cost structure is half the appeal.
