Build Autonomous Agent Fleets
Split substantial work between planners, builders and reviewers. Lightspeed gives each agent a durable session and the tools its role needs.
One agent with a long task list becomes a bottleneck. Substantial work is parallel: someone plans, others build and test, reviewers challenge the result, and a coordinator brings the pieces together.
Lightspeed supports that pattern with agents that delegate to other agents. Each specialist has its own session, instructions and tools, while the parent keeps track of the overall task.
Give specialists a clear job
A profile is a reusable setup for a role. A builder might have repository access and a machine for running tests. A reviewer might read the same source material without permission to change it.
The planner selects an allowed profile and supplies a brief. The child works in its own conversation and returns a result the planner can use. Several specialists can work at once; the planner can wait for their answers or continue with another part of the task.
This division keeps each agent’s context focused. Reviewers can examine the evidence with fresh eyes, and a failed approach can be reassigned without starting the whole project again.
Attach compute where it helps
Some tasks need only APIs and persistent virtual files. Others need a shell, a compiler or a full development environment. Sessions can attach prepared machines for that work.
Separate machines and workspaces let builders try alternatives independently. Shared workspaces let a team collaborate on the same files. The choice follows the task rather than requiring a dedicated running machine for every agent.
Keep the fleet bounded
Delegation has explicit limits: how many agents may be created, how many can work at once, and how long the work may run. Approved profiles determine which specialists a planner can call.
Every session keeps a durable record of its progress. A worker restart can recover that state, allowing a project to continue across tasks that last hours or days. Operators can inspect the delegation tree and intervene as the work unfolds.
A software factory
Consider a feature request that touches an API, a database migration and a user interface. A planner splits it into tasks. Builders implement the changes in prepared environments, testers run the suite, and reviewers examine the diffs and reports.
The planner combines the results, sends incomplete work back, and presents the finished proposal for review. A bot can keep this process available for the next feature, routing each request into its own conversation.
A thousand sessions need not mean a thousand running machines. The fleet can retain its conversations while compute follows the work that actually needs an operating system.
Read next: Deploy Resident Agents for ongoing work, or Accelerate and Automate Migrations for a larger factory. Setup details live in the sub-agent documentation.