You didn't outgrow the agent. You outgrew babysitting it.

You did the hard part already. You found one of the open-source agents, fed it your API keys, pointed it at real work, and watched it go. Most people have never seen that. They have seen a chatbot answer a question. You have seen a thing that holds a goal and acts on it, and for a while that felt like the future arriving a few years early.

Then somewhere the arithmetic flipped. You stopped spending your time deciding what to hand off, and started spending it keeping the agent alive. Dead sessions, keys that needed re-pasting. The update that broke the thing that worked yesterday, with no obvious way back. The slow realisation that you were now responsible for a piece of software you could not read.

Here is the part nobody told you. That was never a failure of your ability. It was the design.

When you self-host an autonomous agent, you do not just get the agent. You inherit two jobs you never applied for.

The first job is security

OpenClaw is roughly 430,000 lines of code. You cannot audit that, and neither can most engineers. When AWS shipped an official one-click OpenClaw blueprint for its Lightsail service this March, a security researcher found it arrived with thirty-one unpatched security updates, including critical holes in the system’s core. People ran it anyway, because the entire appeal was not having to think about that sort of thing. A non-coder running an autonomous agent on their own machine is a sysadmin who was never told they took the job, standing guard at a door they cannot see.

The second job is just math

An agent that is 85% reliable at each step sounds excellent. Chain ten of those steps into one real task and it finishes correctly about a fifth of the time. The failures are rarely dramatic. They compound quietly, because the agent has no real sense of where it is and no way to stop, recover, and pick up where it left off. So you hover. You watch it work in case it wanders off. And the thing you brought in to give you your attention back becomes the single largest claim on it.

You have probably felt the sharp end of this. An agent told to clean up an inbox that deletes most of the inbox instead. It technically did something inbox-shaped. It also did the one thing you would have stopped, if anyone had thought to let you.

That moment is the whole story. Not the capability. The absence of a hand on the brake.

Graduating is not downgrading

So when people talk about moving off the terminal agents, they reach for the wrong word. They call it downgrading, as if the choice is only ever between a powerful tool you babysit and a safe toy that cannot do anything. That choice is false, and it has been false for a while.

The power was never in the command line. The power was in having something that holds a goal and acts on it. You can keep every bit of that. What you can put down is the operating: the patching, the standing guard, the night spent restarting a session so the morning’s work survives. None of that was the autonomy. It was the rent you paid to host the autonomy yourself.

What you actually wanted, underneath all of it, is the oldest arrangement in working life. You tell someone capable what a good outcome looks like and name the few moments you want to be asked about before they act. Then you let them run, trusting they will stop at the edge of anything you would not want done without you. That is not a smaller version of what you had. It is the version that was meant to exist before the software turned you into an operator.

The whole field is quietly circling back to the same answer. The fix for a fragile, unsupervised agent is checkpoints: a system that pauses at the decisions that matter, hands them to a person, and carries on once that person has decided. There is a less technical word for it. Delegation. It is what good directors have always done with capable teams, and it has never required a terminal.

That is the thesis behind be/brief, and it is close to the only learning curve we ask of you. Not a syntax. A habit you probably already have, pointed at a team that can finally act on it. You stay the one who decides. Something capable does the rest. And the agent that tidies your inbox stops, every single time, before it deletes a thing you have not seen.