Compute Workspace
Instances
Running work in the compute layer is an instance you launch, inspect, and terminate. One of them, the scheduler, turns a chat into a recurring, self-improving job.
Instance types
There are two instance types. Both share the same lifecycle (launch, inspect, terminate) but differ in what they run:
| Type | What it runs | Lifecycle |
|---|---|---|
| GPU | Launched compute for training or inference against a GPU offer. | On demand. You launch and terminate it. |
| Scheduler | A chat that re-runs on a cron cadence. | Recurring. It re-arms itself after every run. |
GPU instances have their own page under GPU offers. The rest of this page is the scheduler.
The scheduler
A scheduler instance is one scheduled_chat: a prompt, a cron_expression and timezone, an optional capability chain, and the id of the chat it belongs to. On each due tick the runtime claims the instance, runs it, records the result, and computes the next run, so the loop keeps running on cadence with no manual re-trigger.
Because the instance is bound to a chat_session_id, every run appears back in that same chat as a normal turn. Automating it does not hide it. The runs, and anything they surface, land where you can read them.
The recursive schedule loop
A scheduler instance suits work that should keep improving something over time. A typical loop preps and cleans recent data, snapshots it, evaluates your production prompts against that snapshot, and recommends improvements, then re-arms for the next run.
Every Monday 9am, clean last week's traces,
snapshot them, evaluate my production prompts,
and recommend improvements.Cases that suit a recursive schedule
- Continuous data prep and cleaning of the last period's traces.
- A weekly regression check of production prompts against a fresh snapshot.
- Recurring prompt-improvement recommendations from evaluation results.
- A scheduled cost and usage digest per project.
Scoring for the evaluation step is covered under Evaluation.
Setting a schedule from the chat agent
Schedules are created conversationally, not from a config screen. You describe the cadence in chat; the agent turns it into a schedule, validates the cron cadence before saving (a malformed expression self-corrects rather than failing silently at run time), and asks you to confirm the plain-English summary. Nothing schedules without that confirmation.
You: Every weekday at 9am, clean yesterday's traces and evaluate my prompts.
Agent: Here's the schedule I'll create.
cadence: weekdays 09:00 (Europe/London)
steps: clean, snapshot, evaluate, recommend
Confirm to schedule it?
You: Confirm.The schedule is saved against the chat it was created in, so every subsequent run, and any recommendations it produces, returns to that thread. Recommendations are never applied automatically. Promotion stays a human decision (see Evaluation).