Fork into independent branches.
Each branch is a full sub-chain — its own Investigate, Plan, Create — on its own context window and model.
Run 20 agents on one task and watch every one at once — live tiles on a bird's‑eye canvas, not buried in text logs. Fan out, stream, converge into a single result.
A Parallel node forks one prompt into twenty branches. Each runs its own chain and streams its state onto its own tile. A Merge node converges them back into one result. Three steps:
Each branch is a full sub-chain — its own Investigate, Plan, Create — on its own context window and model.
Each agent reports running, paused, done, or error straight onto its tile — no hidden buffer. The whole swarm is legible at a glance.
An Evaluate pass reconciles every branch — dedup'd, ranked — into one result, ready for Review or Apply.
Hit Run swarm. Tiles light up as agents stream, turn green as they finish, then converge into one result — the same live surface that ships in DISPATCH.
Twenty agents is only useful if you can tell, in one glance, whether the swarm is healthy. Agentic Streaming makes parallelism observable: state lives on the canvas, not in a log you scroll after the fact.
Branch, retry, cap concurrency, loop, nest, gate, timeout — the controls under the swarm.
open TestingLay out goals, phases, and milestones right on the graph the swarm runs against.
open ActiveWatch the same overhead canvas together — your whole team on one running swarm.
open