Cloud Agents Platform
A harness for agents that outlive the request.
LeelE is a durable, event-driven harness for long-lived AI agents, designed to run on infrastructure you own — so work checkpoints as it happens, destructive actions wait for your approval, and every token is accounted for.
Being built in the open, one proven layer at a time. The contract layer and state machines are live and inspectable; the executor that runs them is not built yet. Every capability below carries its real status.
- plancheckpoint 1
- search:webcheckpoint 2
- read:sourcescheckpoint 3
- redeployprocess replaced
- resumefrom checkpoint 3
- synthesiserunning · 4.2s
- publish:reportawaiting approval
An illustration of the execution model: three checkpoints, a redeploy that replaces the process mid-flight, and a resume from the last checkpoint rather than a restart. The contracts behind it are built; the executor that runs them is not yet.
What it does
Everything a long-running agent needs to be trusted with real work.
Durable runs
in progressWork is designed to checkpoint as it happens, so a run survives a deploy, a crash or a wait for approval — and resumes from where it stopped rather than starting over. The state machine and contracts are built; the executor is next.
Memory that lasts
plannedSessions will keep their history, recall what matters from earlier, and compact the rest, so context stays useful well past the point a window would overflow.
Approvals before actions
in progressDestructive tools stop and ask. The approval contract binds each confirmation to the exact call, policy and capability envelope it was granted under, and is signed so it cannot be forged or replayed.
Costs you can see
plannedEvery run will be attributed by user, model and token type — including cached tokens — so spend is measured rather than estimated.
Parallel agents
in progressFocused agents run alongside the main one with constrained tools, reporting back through a mailbox rather than flooding the conversation. The authority model that decides who may stop whom is built.
Yours to run
plannedDesigned to be self-hosted on infrastructure you control, with your own model provider, so no data leaves your account by default. Today it runs only as this deployment.
How it works
Two execution planes, because a chat reply and an overnight job are not the same problem.
Most platforms pick one and make the other awkward. Interactive turns take the direct path; durable work checkpoints so it can be interrupted without loss.
DesignedThis is the settled execution model, not a description of running software — the executor is the next cycle.
Answers while you wait
An interactive turn is designed to run in-request and stream straight back — no queue, no scheduler, no indirection, because the fastest path is the direct one.
- Will stream token by token
- Will resume if the connection drops
- Bounded by an explicit time budget
Work that outlives the request
Long and scheduled work is designed to checkpoint at every recovery boundary. Between turns nothing runs, so a job can wait hours for a subtask or days for an approval and still be one run.
- Will survive deploys and restarts
- Will retry a step, never the whole run
- Suspends free while awaiting approval
Access
Built in the open, deliberately.
LeelE ships in small, verified increments. Access is granted by invitation while the platform is early — not to manufacture scarcity, but because every account is one we can actually support.