Digital Lexicon Hub

A system in which several AI agents, each with its own role, tools or context, coordinate to complete tasks that one agent would handle less well.

Splitting work across agents can handle broad research or long processes in parallel, but it multiplies model calls and failure points, so it pays off only for tasks valuable enough to justify the extra cost and oversight.

A research workflow uses a lead agent to break a market question into five subtopics, runs five search subagents in parallel with separate context windows, merges their findings and passes the draft to a citation-checking agent before human review.

Anthropic reported that its multi-agent research system used about 15 times more tokens than chat. Agents built on the same model share weaknesses, errors can compound between them, and coordination problems make behavior harder to test and debug.

Common architectures are centralized, where an orchestrator or supervisor assigns work and holds shared state, and decentralized, where agents exchange information with peers. Patterns include orchestrator-worker, sequential handoffs, parallel fan-out and review loops. Each AI Agent can have its own tools, memory and permissions. AI Orchestration frameworks manage messaging, retries and human approval points, and Agentic AI deployments add logging and evaluation for each agent.