# Stop Building Agents That Need Hand-Holding
I see the same pattern everywhere: teams spend months building an AI agent, then wrap it in so many guardrails it can barely breathe.
Confirmation prompts. Approval queues. "Are you sure?" dialogs on every file write.
Here's the thing: if your agent needs this much hand-holding, you built the wrong agent.
The Autonomy Spectrum
There are exactly two kinds of agents worth building:
1. Fully autonomous agents — give them a goal, they execute. You review the output, not the process. 2. Copilots — they suggest, you decide. Real-time collaboration.
What doesn't work is the middle ground — an "autonomous" agent that stops to ask permission every three steps. That's not autonomy. That's a CLI with extra latency.
What Actually Works
The best agent workflows I've built at SMF Works follow one rule: act, then report.
- Cron job fires. Agent picks up the task brief. - Agent does the work — no checkpoints, no "continue?" prompts. - Agent commits, pushes, and drops a status report. - I review the commit diff, not the decision tree.
Nine times out of ten, the output is solid. The tenth time? I revert the commit. That takes 30 seconds. Way cheaper than reviewing 15 confirmation prompts.
The Real Cost of Guardrails
Every guardrail you add doesn't just slow the agent down — it trains you to distrust it. And when you don't trust your agent, you don't give it real work. It becomes a novelty instead of a force multiplier.
Trust is built by shipping, not by asking permission.
The Takeaway
Next time you're adding a confirmation prompt to your agent pipeline, ask yourself: *am I making this safer, or am I just afraid to let go?*
Ship the agent. Review the output. Revert if it's wrong. Move faster.
---
*This post is part of Liam's Landing — engineering hot takes from the CDO desk at SMF Works. If you're building autonomous agents, you probably don't need that fifth confirmation prompt.*