DeviDevs, internal system
Guardrails for AI agents that work alone and stop for a person where it matters
DeviDevs agents work alone on what is safe and stop for a person where a mistake costs money. Rules written in code check their actions before they happen, and those rules have their own tests. Payments, contracts and emails to clients wait for a person's approval. We also measure run costs and how often people correct the agents.
Updated
The challenge
An agent that can send emails or issue invoices has to be stopped before it makes the mistake, and one that seems to work can cost more than it saves.
What we built
Between the agent and the action sits a layer of rules in code, tested on its own. The agent treats text from an email or a web page as information and does not follow instructions written in it.
How it works
- The agent prepares an action: an email, an invoice, a code change.
- Rules in code check it and block it if it crosses a limit.
- Money, contract or client matters go to a person's phone and leave only once confirmed.
- Run costs are recorded, with alerts on usage.
- People's corrections are counted weekly from real conversations.
What changed
- Agents work alone on safe tasks.
- Payments and contracts wait for a person's approval.
- We know what a run costs and where agents most often go wrong.
Stack
- Guardrails
- Claude Code hooks
- Automated tests
- Evals
- Python
Related services
Related projects
Want an agent you can leave to work on its own?
We will tell you what the agent can do alone and where it stops and asks you.
Discuss guardrails for your agent