Automations / Development
Terminal-Bench 89% (Opus 5.5)
The exact agent that scored 79/89 (88.8%) on Terminal-Bench 2.1, above Claude Code (78) and Codex CLI (72) on the same model. Claude Opus 5.5 at high thinking, a 300-byte system prompt, read + bash only, Opus 4.8 fallback on refusals. Give it a terminal task; it works until it is verified done.
01
The agent brief
Before you finish go over each statement the user task has requested you to do.
02
What the agent does
Describe the task to complete.
FAQ
Common questions
What does the Terminal-Bench 89% (Opus 5.5) automation do?
The exact agent that scored 79/89 (88.8%) on Terminal-Bench 2.1, above Claude Code (78) and Codex CLI (72) on the same model. Claude Opus 5.5 at high thinking, a 300-byte system prompt, read + bash only, Opus 4.8 fallback on refusals. Give it a terminal task; it works until it is verified done.
Do I need an API key or integration?
Usually not. The agent works in your own logged-in browser and on your own computer or cloud VM, the same way you would, so tools without an API work too.
Will it act without asking me?
You see the plan and can stop or redirect the agent at any point in the chat. Anything public, like posting or sending, waits for your approval unless you tell it otherwise.
Can I change the instructions?
Yes. Running it opens the brief in the composer, so you can edit the task, add your context and pick the model before it starts.
What does it cost?
The free plan runs it on Claude Sonnet. Paid plans start at $20 a month and include every model provider.
Hand this one to an agent.
Free plan, your own cloud VM, no card.