Skip to content

Use case

Competitor research agents that report back and shut down

Send several agents to research competitors in parallel, each in its own sandbox, under one dollar budget, and get a single brief with a receipt for how it was made.

"Research my top ten competitors" is a good job for agents: wide, parallel and easy to check. It is also easy to overspend on, because the natural way to go faster is to add more agents.

How does it run?

You describe the goal and a budget, for example $6, and start a fleet. Each agent gets its own sandbox and its own short-lived credentials. They read the web through the gateway, which is the only way out of the sandbox, and every model call spends from the shared budget through leases.

When the work is done, the fleet returns one brief and the sandboxes are destroyed. The trace stays.

Why one budget for the whole fleet?

Because the question you care about is "what did this research cost", not "what did each worker cost". A fleet-level cap means scaling out is safe: more agents finish sooner, but they draw from the same $6.

What do I get at the end?

The brief, plus a receipt: which agents ran, which model and version, which policy, what was spent, and the trace that led to each finding. If a claim looks odd, you can see where it came from.

Updated

Questions

How many agents can work on one research task?
A fleet can run one agent or several under the same manifest and the same budget. The budget caps the whole fleet, so adding agents cannot raise the spend.
Do research agents keep running afterwards?
No. A one-off research run starts its sandboxes, returns the result and cleans up. The trace and receipt are kept.

A new home for autonomous work.

Big ambitions. Steady hands.

Bring your first agent aboard. Heighliner is in early access, and we’re building with our first crews.