Orchestrator + Subagents + Synthesis
The full multi-agent research architecture animated.
This animation plays the full life of one research query through Anthropic's multi-agent architecture.
Act 1 — Plan. A query arrives at the lead agent, which writes a plan into persistent memory (storage outside its context window). Watch the plan survive the "context wipe" flash — that persistence is why a truncated lead agent can recover instead of starting over.
Act 2 — Fan out. The lead spawns three to four subagent lanes at once. Each lane runs its own search-and-read loop with its own context window, shown as a token budget bar filling as it works. Notice that every subagent also fires several tool calls in parallel — parallelism at two levels is what cuts research time by up to 90%.
Act 3 — Synthesize. Each lane returns a thin condensed-summary edge — visibly thinner than the fat token streams it consumed. That compression is the point: the lead reads summaries, not transcripts. The findings merge in a synthesis node, pass through a citation step that stamps sources, and exit as one answer.
Watch for the failure case: one subagent lane dies mid-run, and the lead compensates from the surviving lanes plus the saved plan. Also watch the running token counter comparing 1x chat, 4x single agent, and 15x multi-agent — the cost that makes this architecture a deliberate choice, not a default.