Designing Agent Fleets

What you’ll learn and how you can apply it

• Design an agent fleet as a systems architecture: Define the workload, run contract, worker contract, state boundaries, failure boundaries, and control-plane responsibilities
• Build around the run as the unit of work: Model each agent run as a durable, verifiable task with a workspace, artifacts, events, an exit condition, and a verifier
• Design the fleet control plane: Coordinate queues, worker capacity, leases, scheduling, durable state, verification, failure handling, and spend limits

Course description

If your product runs general-purpose agents in cloud sandboxes, or your team has adopted coding agents, you’re already operating a fleet of autonomous agents running concurrently in isolated environments, producing work someone has to trust. Managed agent runtimes, Kubernetes Agent Sandbox, and self-hosted operators show the same pattern emerging across cloud and open source systems. The hard part is the control plane that decides what runs, where it runs, whether the work holds up, and when to stop.

Join Sajal Sharma to explore that layer as a systems design discipline. The architecture follows from the agent worker: the run as a durable, verifiable unit of work; the contract between workers and the control plane; and the queue, scheduler, state, and acceptance loop required to coordinate a pool. Essential operational signals appear where they affect the design: queue depth, worker health, failed runs, verifier results, and spend. Concepts are folded into a systems design case study: define the workload, assign state and failure boundaries, assemble the reference architecture, and map it to open source and AWS components. In just two hours, you’ll design and size an agent fleet, and you’ll be equipped to explain its failure modes and evaluate platforms that provide pieces of a fleet.