I fired my AI employee. Here's what I lost.
What the generalist agent cost me
$100 in 2 days from simple prompting and heartbeats. OpenClaw struggled as a generalist — it underperformed with cheaper models (although Haiku was much better than the OpenAI alternatives), it intentionally bloated the context window with excess information, and it couldn't route to the appropriate model effectively.
512 security vulnerabilities. 8 critical. Roughly 1 in 4 community-built skills had at least one vulnerability. The most popular skill on its marketplace was literally malware.
Constant framework changes. With the explosion in popularity and its inherent "use the framework to improve the framework" approach, OpenClaw received a flood of contributions and underwent more change than just the name. Shifts in even the basic interfaces exposed the immaturity of the tool. Nearly every blog post and tutorial I found was written for a version that no longer existed.
What Peter and the open source team have built is impressive, and it taught me a lot about agents. But the world isn't ready for the generalist AI assistant yet. The models aren't cheap enough. The context isn't big enough. The frameworks aren't mature enough. And most importantly — the safety features don't exist to allow unbridled access to all of your systems and data, especially when bad actors are specifically targeting frameworks like this.
Some people are using OpenClaw effectively, but they're treating it as an agentic workflow tool — not a generalist assistant — and there are honestly better frameworks available for that.
So I rebuilt
Agentic workflows the right way. Proper orchestration layer. Scheduled triggers instead of always-on token burn. Smart model routing. Prompt caching. Human-in-the-loop before anything touches my outbox.
The monthly cost is reasonable, and the new system does exponentially more to manage my business than the old one ever did.
The real problem wasn't the technology
I made the same mistake managing my AI agent that I've made with new hires.
Not enough direction. Not enough onboarding. Unclear metrics and goals.
With people, that's bad management. With agents, it's a death spiral. The difference is humans fill in the blanks. They ask questions, read between the lines, make reasonable assumptions. Agents don't. They just... go. Confidently. In the wrong direction.
I gave one of my agents a broad project with loose instructions. A person would have come back with clarifying questions along with their own inspiration. The agent came back with 47 minutes of token burn and a deliverable I couldn't use.
The fixes
The same ones I'd give any manager struggling with a new hire:
Specific deliverables, not vague responsibilities. "Break this project into tasks with owners, deadlines, and dependencies" — not "manage this project."
Written standards, not tacit knowledge. If it's not documented precisely enough for a machine to follow, it wasn't documented well enough for a person either.
Check-ins early and often. Daily reviews and pair programming. The same cadence that works during an employee's first 90 days.
What actually moved the needle
One thing that made a real difference: planning the work with the agent before having it start. Defining the scope, breaking down the steps, agreeing on the output format — then executing. When I treated the agent like a collaborator in the planning phase instead of just handing it a task, the quality of the work jumped immediately.
The lesson
The agents didn't expose a technology problem. They exposed a management problem. Most failures traced back to something I could have managed better — and agents just made it obvious faster because they don't compensate for unclear instructions.
What makes you a strong manager of people will make you a strong manager of AI. And what you're not managing well with people, agents will make painfully obvious.