Agent Cloud: run and verify agent tasks

What is Agent Cloud?

Agent Cloud runs a coding or testing agent on a disposable CPU VM and checks the result instead of trusting the agent's own claim that it is done. There are two ways in, with different behaviour: badgr launch for a quick run, and tracked agent jobs when you need a record you can poll, cancel and look up.

  • Coding agents: cline, claude, codex, opencode, commandcode
  • Browser and end-to-end testing with playwright
  • A bare task: badgr launch "<task>" resolves a repo and an acceptance command for you

Launch your first agent

npm install -g badgr-cli
badgr login
badgr launch cline "Fix the checkout bug"

cline uses Badgr-provided model access, so there is no account to connect. claude and codex ask you to connect your own account the first time. playwright needs no model account.

What happens

  1. 1.Badgr starts a disposable CPU VM. No --max-cost is needed for the built-in agents.
  2. 2.The agent runs your task against the repo.
  3. 3.Badgr runs the acceptance command and reports Verified or NOT VERIFIED.
  4. 4.The VM is torn down. Pull the result with badgr pull or badgr artifacts.

Cost control

  • Agent VMs are CPU machines, billed while they run and stopped when the task ends.
  • Model usage for claude and codex is billed by your own account. Badgr only provides the VM.
  • badgr down <id> stops a run, and badgr status shows what is running.

Get results

  • badgr pull <id> brings a code-editing agent's patch back as a local git diff or branch.
  • badgr artifacts <id> downloads test reports, screenshots and traces.
  • Tracked jobs (badgr job or POST /v1/jobs with type: "agent") return a job_id you can poll, and store a receipt.

What this does not cover

  • Agent Cloud is a runner and verifier, not a hosted agent product. You choose the agent and own its prompts and credentials.
  • Real-time log streaming is polling every few seconds, not true streaming.

Next steps