Software factories
Route agent workflows to the right agent and model. One call.
Your factory controls the queue, policy, and budgets. Point any workflow step to the AgentSky API — it runs the agent, model, isolated sandbox, persistent state, and event stream on the same call.
Tycoon
HeyBoss AI
WebJourneyAny harness, one API
Claude Code, Codex, Hermes, OpenClaw, pi, and every other supported harness are behind the same API contract. Switching the agent for a workflow is a one-word change.
Budgets and policy
Attach a hard spend cap to each session at creation. Set per-tenant billing routing through metadata. AgentSky meters and stops when the cap is reached.
Parallel execution
Each workflow step runs in its own isolated cloud computer. Your orchestrator controls concurrency; AgentSky accepts each job and executes it independently.
One clear boundary
Your factory routes. AgentSky runs.
Keep the workflow graph in the system that already understands your product. Call AgentSky whenever a step needs an agent with a real computer and durable state.
Queue + concurrency
Your factory: Priority, fan-out, global limits, and backpressure.
AgentSky: Accept asynchronous turns; provision sessions eagerly or on first work.
Tenant isolation
Your factory: Map customers, roles, repositories, and policy to session metadata.
AgentSky: Scope API access by universe; isolate each session in its own workspace.
Budgets + cost
Your factory: Set limits by tenant, workflow, team, or job.
AgentSky: Meter runtime usage; stop new work when the prepaid balance is exhausted.
Audit + observability
Your factory: Record routing decisions, approvals, and business outcomes.
AgentSky: Expose session metadata, live events, message history, and runtime logs.
Reliability + recovery
Your factory: Decide job retries, fallbacks, and terminal failure policy.
AgentSky: Preserve the runtime volume across restarts; restore context from backups after machine loss.
By the numbers
- 150,000
- 250,000
- 50,000+
- $0
Read the Tycoon story →“Tycoon’s AI CEO routes every task to the right agent and model—Claude Code, Codex, Hermes, OpenClaw, pi—through one API. Adding an agent to the roster is a parameter, not an integration. Each harness and model combination is one call; the runtime for all of them is the same.”
Compare approaches
Build it yourself, or use the runtime.
| Capability | DIY runtime | Agent SDK | AgentSky |
|---|---|---|---|
| Agent provisioning | Build and maintain per-harness process lifecycle | One SDK, still self-hosted | Fully managed, cloud computer per session |
| Session state | Design and persist your own context store | SDK manages state locally | Persistent workspace, recovery, event history |
| Streaming events | Each harness has a different event format | Normalized per-SDK | One SSE contract for every harness |
| Multi-harness support | Re-implement per harness | Multiple SDKs required | One API; switch is a one-word change |
| Compute + billing | Provision and pay vendors separately | Your responsibility | Metered per second; one balance |
Your software factory. One cloud-agent runtime.
Tell us how your orchestrator works. We'll show you where AgentSky fits in.