Why Every Agent Needs a Kill Switch (And What We Built)
You're deploying an agent so it can do things without asking. That's the point.
That's also the problem. When it does something you didn't intend—how fast can you stop it?
The Labs Say So
Anthropic: "The difficulty of maintaining human oversight as systems become more autonomous." OpenAI's GPT-4 card: "Actions with real-world consequences that are difficult to reverse." Stanford HAI: "Control and oversight mechanisms" are the gap.
Not theory. That's why every lab is pouring money into safety. You're the one putting an agent on a server. You need the controls.
What Actually Goes Wrong
Your agent can:
- Act without asking — You want that. Until you don't.
- Call any API — Useful. Also dangerous. Who decides which domains?
- Run when you're not there — 24/7 is the dream. Also: runaway costs, bad prompts, data you didn't mean to touch.
One bad prompt. One compromised skill. One exposed dashboard. You need to stop it now, not "when it finishes."
Three Controls That Matter
We built Clawctl around three things. Not features for the brochure—things that unblock deployment.
1. Instant Stop
Not "gracefully finish." Not "queue shutdown." Stop.
Click the Pause button in the dashboard. Agent halts. When you're ready, click the Resume button.
One click. From anywhere.
2. Network Allowlists
By default your agent can only hit:
- api.anthropic.com, api.openai.com
- github.com, registry.npmjs.org
- Domains you explicitly add
Everything else: blocked. At the network layer. Not "we log it"—we block it.
Add approved domains and view egress stats from the Network tab in your dashboard.
3. Full Audit Trail
Every prompt. Every tool call. Every API hit. Every file touch. Timestamps. Searchable. Exportable.
Something weird happens? You replay exactly what the agent did. Security asks questions? You have answers.
Why This Unblocks You
The goal isn't to make agents less useful. It's to make them safe enough that you can actually deploy them.
With a real kill switch, clear egress, and a log of everything—you can run it overnight. You can give it production access. You can tell security "we have controls."
Without that? You're one bad prompt away from explaining to your CTO why the agent did that.
Every Clawctl plan: kill switch, egress allowlists, audit logs. Not enterprise upsells—table stakes for running agents in prod.