One agent is still one worker#
However good your coding agent is, it works the way you do: one task at a time, in one checkout, inside one context window. Give it three bugs and it takes them in order. Give it a big refactor and the context fills up with the parts it has already touched. The bottleneck stopped being how well the agent writes code a while ago. It is that there is one of it.
Delegation is a tool call#
Connect your agent to Rock8Cloud over MCP and it gains task_agent. It describes a task, and Rock8Cloud starts a cloud agent in its own sandbox with its own checkout of the repository and its own context window.
The important part is what the call returns: a session, not a finished job. Your main agent is not blocked waiting. It hands off the work and carries on, which is what turns one worker into a pool.
Work comes back one of two ways. A research agent reads and reports, so what returns to your main agent is information: where a thing lives, why it breaks, what depends on it. A coder agent changes the checkout, so what returns is a branch and a pull request. You choose which shape you want when you hand off the task, and the two compose - findings from the first can brief the second.
Four ways to fan out#
Independent tasks, at the same time. Three bug reports become three sessions and three pull requests. They do not queue behind each other because they never shared a workspace.
Scout, then build. Send a research agent in read-only to map an unfamiliar part of the codebase and report back. Hand its findings to a coder agent as a briefing, so the expensive session starts already oriented.
The right model for each job. list_agent_models shows what your organization can run. Route the mechanical work to a cheap model and save the heavyweight one for the task that actually needs it, per session rather than per account.
Work that outlives the session. Tasks run while you are in a meeting and while you are asleep. A backlog of small fixes can be worked overnight and be waiting as pull requests in the morning.
Your agent becomes the supervisor#
Fan-out is only useful if something collects the results. list_agent_sessions and get_agent_session let the main agent keep track of what it started, check which runs finished, and report back in one place. You ask one question and get one answer, with several sandboxes behind it.
Throughput goes up, risk does not#
Every coder session ends the same way it would have if you had done the work: a branch and a pull request, with a preview environment and a review. Nothing merges itself, and nothing an agent does reaches production because a different agent asked it to. You are multiplying the work that arrives at your gate, not moving the gate.
Start free#
Give your agent a team. 60 days free, no credit card required.