Adopting an agent platform is a decision about where your code runs, which models see it, and who controls the bill. Here is how each of those works, in plain terms.
You are not locked into platform-billed inference. Connect your own provider keys and model subscriptions, and route spend to your allowances where you configure it. Usage attributes every run to the account that paid for it.
If you would rather not manage keys at all, stay on platform credits. Most teams do both: platform credits for exploration, their own keys for volume.
Token discipline is built into the harness rather than left to whoever wrote the prompt. Context is scoped per agent, routing decides which model is worth using for which stage, and caps apply before spend happens.
Use a strong model where judgement matters and a cheaper one for mechanical stages.
Set a ceiling per organisation. Runs stop rather than quietly continuing past it.
Every run carries a cost ledger with a token breakdown by stage, plus the account it was billed to.
Runs execute in isolated cloud sandboxes on pooled runners, scoped to one organisation. Secrets are managed and scoped to the run that needs them rather than baked into an image, and repository access follows the scope you granted the GitHub app.
We run the infrastructure, the runner pool and the upgrades. Sign up and connect a repository in minutes.
For teams whose code cannot leave their estate: a sandboxed deployment inside your own environment, with your providers and your network rules.
Agents open pull requests; humans merge them. Where something is still being built, we will tell you rather than put it on a feature list. Ask us directly about anything on this page and you will get a straight answer.