Blog
Meet Multi Agents: running long context tasks without losing the thread
Today we’re introducing Multi Agents.
Most engineering tasks don’t fail on model capability. They fail on context.
A single agent has to plan, execute, inspect changes, recover from mistakes, and keep the whole task coherent inside one context window. That works for smaller tasks. But on larger codebases and more complex problems, the context window becomes the bottleneck.
Large tasks either overflow the context window or create enough context pressure that the agent starts losing the thread.
Multi Agents changes that execution model.
Multi now runs a supervisor agent that plans the work and delegates focused steps to subagents. Each subagent operates in its own context window, so individual steps stay scoped instead of competing for the same context budget.
The supervisor keeps track of the overall direction. Subagents handle individual pieces of the task. You stay in the loop with visibility into what each agent is doing.
More agents without visibility is just more chaos. Multi Agents are built around supervised execution.
Use them for the kinds of tasks where one agent usually starts losing the thread: repository refactors, multi file migrations, long-running investigations, increasing test coverage, and multi step feature work.
When to use Multi Agents
Use Agents when the work is large enough to benefit from decomposition:
- repository-wide refactors touching many files
- complex features across multiple subsystems
- staged migrations, like schema changes, query updates, and tests
- long-running investigations
- test coverage work across a codebase
- tasks where parallel branches can reduce serial bottlenecks
Keep it off for quick fixes, single-file edits, or speed-first prompts. Agents mode adds orchestration overhead, and simple tasks are often faster in single-agent mode.
How it works
- Enable Agents before starting the task from the Agents toggle in the action bar or the
/command menu. - The supervisor receives your request, breaks it into smaller steps, and spawns subagents.
- Subagents execute independently and report back.
- The supervisor coordinates the work and returns a consolidated result when the task is complete.
You stay in charge
Throughout execution, you can:
- Monitor workers inline and in the Subagents View
- Open subagent details to inspect progress, outputs, and usage
- Respond to actionable items (approvals, questions, or max-step continuation)
- Track itemized token spend in the Usage View with supervisor, subagents, and total token breakdowns
Try Multi Agents
Give Multi Agents a try on your large task.
We’d love to hear how it feels.
