Agent fleet
Run an agent fleet without building one.
AgentSky creates one isolated cloud agent per task through a single API call. Each sandbox starts independently, runs in parallel, and streams structured results — the same approach behind 40,835 tasks run across 10 agent stacks. You keep queue and output policy; AgentSky operates the sandboxes.
Running many agent tasks in parallel means operating infrastructure teams never signed up to build.
Choose AgentSky when
Product teams starting many isolated agent tasks through one API while retaining control of task routing, budgets, and downstream outputs.
Choose another approach when
Own the execution layer when you need self-hosting, custom queue placement, GPU training, kernel-level controls, or workload guarantees outside AgentSky's contract.
How it works
One job. One complete cloud agent.
One API call, one isolated agent per task
Create sessions programmatically — each task gets its own cloud sandbox, isolated from every other run. Priced per session, not per seat or per machine.
Right agent stack routes to each task
Route different task types to different harness and model combinations. The API contract stays the same regardless of what runs behind it — the fleet is heterogeneous by default.
Structured events stream back per session
Stream events per session. Artifacts, logs, and history land per run — consistent format across every agent in the fleet, ready to aggregate or route downstream.
You choose
- Task list and instructions
- Agent stack per task
- Concurrency and budget
AgentSky runs
- A cloud sandbox per session
- Streaming events
- Lifecycle and recovery across the fleet
Typical stack: Hermes + DeepSeek V4 Flash
See the agentSoftware factories
AgentSky's software factory pattern creates one isolated cloud agent per task through the Agent API. The product owns queue logic and output routing; AgentSky operates every sandbox, event stream, and recovery layer.
Evidence scope: A reference architecture, not a capacity or throughput guarantee; concurrency, budget, and workload fit still require validation.
Read the reference architecture