You build the ordering bot, the scheduling agent, the inventory checker. It works great in the demo. Then it's live on a real client's business — and you're the one who finds out something went wrong when they call you, not when a dashboard tells you first.
A timeout, a bug, a loop — the same payment goes through twice before anyone notices.
An action or endpoint it was never supposed to touch gets called anyway, quietly.
Injected instructions in a scraped page or crafted input try to override its constraints.
Works alongside whatever you've already built. No agent framework lock-in.
What actions, what endpoints, what rate limits, what payment rules. One JSON file per agent. Use the config generator to build it in a couple of minutes instead of hand-writing it.
Your agent's code sends each intended action to a lightweight Cloudflare Worker, which checks it against the manifest and answers allowed or blocked. It runs on your own Cloudflare account.
A live feed of what the agent's actually been doing, and a suspend switch that blocks its very next request.
Worth knowing: WarilyWatch only checks actions your agent asks about. It can't stop an agent that skips the check, and the phrase filter is a tripwire, not a guarantee. Think safety net, not sandbox.
Full Worker + dashboard code. Deploy it yourself and run it on your own Cloudflare account. This is the whole product, not a crippled trial. Cloudflare's free plan has small daily limits (around 1,000 KV writes a day), so plan on their $5/month Workers Paid plan for real traffic.
Planned: we host the KV store, handle alert delivery, and keep the dashboard updated. Bill your client $15–20/mo and pocket the difference instead of managing infrastructure yourself.
WarilyWatch isn't a security company's enterprise product wearing a friendly logo. It's built by a solo dev who ships small AI-powered apps for small businesses — same problem you're solving, same stakes, same "it's 11pm and I'd like this to just work" energy.