Multi-Agent Orchestration with Fusion

Fusion is one of Atomic Agent's three run modes, next to Local and Cloud. It gives one job to two models: an orchestrator that plans, writes a brief for each part and reviews the result, and workers that build those parts, several at once. The orchestrator never edits files itself; it checks the work and sends weak parts back out.

Multi-agent orchestration means one coordinating agent splits a task, hands the pieces to other agents and assembles what comes back. In Fusion the coordinator is the orchestrator model, the workers are short-lived agent sessions that write their files straight to disk, and the orchestrator can check each part with syntax checks, test commands or a headless browser before it accepts it.

Both work, and you can swap the two seats at any time. A cloud model planning for local workers is the default and suits jobs that split into many small parts. A local model planning for cloud workers suits code-heavy jobs with a lot of output, because the long writing happens on the cloud side.

Not always. Local workers have no token bill, so with cloud planning and local building only the orchestrator spends API tokens. Long review rounds can still cost more than a cloud model working alone, so set a money cap per turn (llm.runMode.fusion.turnBudgetUsd in the config file): once it is reached, the turn stops starting new work and reports back.

A 27B-class model at 4-bit is practical from a GPU with around 24 GB of memory, or an Apple Silicon Mac with enough unified memory. Parallel local workers share one GPU, so on smaller cards a run can stretch into hours. In the swapped setup the local model plans, so it needs the same hardware. With both seats in the cloud, any machine that runs the app is enough.

Only the local side does. Whichever seat runs in the cloud sees your request and the file contents it works with, so that provider receives them. If nothing should reach a cloud provider, use Local mode instead of Fusion.

In the desktop app, pick Fusion in the composer and choose a provider for each role; a local seat also lets you pick which downloaded model runs. In the terminal version, run /runmode fusion, and /runmode swap to trade the seats. You need two models, usually a cloud model with an API key and a local model. In the terminal they must come from two configured providers.