serverless agents
Your agent is
just code.
An agent is a directory. Instructions, tools, connections, and config, versioned in your repo and reviewed like any other code. Deploy it and it runs serverless on OpenComputer: hibernating between invocations, resuming with its state intact.
agents as code
Every part of the agent has a place.
Nothing lives in a dashboard. The identity, the instructions, the tools, even the eval suite ship in the same directory.
the serverless part
Between invocations, your agent costs nothing.
Every agent gets a real machine, not a stateless function. We just make the machine disappear when it's idle.
01 · invoke
Wake
A trigger fires. The agent's machine wakes in milliseconds with memory, workspace, and context intact.
02 · run
Work
The agent thinks, calls its tools, writes to its workspace. Long tasks are fine; it's a real computer.
03 · checkpoint
Snapshot
When the turn ends, the whole machine is checkpointed: process, memory, filesystem.
04 · sleep
Sleep
It hibernates for free and resumes exactly where it left off. Next week's run remembers last week's.
you pay for seconds of compute, not for an agent that exists
quickstart
Build your first serverless agent.
The quickstart walks you through gmail-summarizer, an agent that reads your inbox and writes you a digest. From empty directory to a deployed, scheduled agent.