Agents now answer customers, issue refunds, write code and change records in live systems. Most teams still pick one by reading the README, checking the GitHub stars and hoping.
We build agents ourselves, and a demo never told us enough. A demo shows an agent doing its job once. It doesn't show what happens when a web page the agent reads has hidden instructions in it, or when the agent has far more access than the job needs.
Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, and names inadequate risk controls as one of three reasons. We wanted a simple way to see those risks before an agent ships.
What makes an AI agent trustworthy
We wanted a way to check our own agents that anyone else could use too. That meant four things:
- Anyone can check an agent by pasting a link, with no sales call and no account.
- The rating looks at how the agent is built, such as its limits, approvals and tests. Popularity doesn't count.
- Every rating points to the code or the test run behind it.
- A safeguard someone has seen working counts for more than one described in a README.
So we built AgentStripes to do those four things.
How we check AI agent security and guardrails
AgentStripes gives an AI agent from 0 to 7 stripes. It checks 40 capabilities in 12 areas, from security and human oversight to observability and runtime governance.
Some capabilities cover AI agent security, such as keeping secrets out of the code, keeping untrusted input apart from instructions to make prompt injection harder, and giving the agent no more access than it needs. Others are guardrails on what the agent does, such as a limit on every run and an approval step before it changes anything real.
Each capability is rated on how well it's done and how strongly that's proven. A safeguard read in the docs earns less than one seen in the code, and one seen holding up in a test earns the most. A rule that exists only in the system prompt earns little, since the right input can get the model to ignore it.
A few serious gaps hold an agent down whatever its score. An API key committed to the repository means 0 stripes. Stripes 1 and 2 can be earned from the code. Above that, the safety behaviours have to be shown working, first in the agent's own runs, then in a test anyone can re-run or in our sandbox, and for stripe 7 by an independent tester.
The full method, with the checklist and a worked example, is in our guide: How to Evaluate an AI Agent Before You Deploy It.
Why the rating has to be hard to game
A trust rating only helps if people believe it. The rules are published, and they apply to Algen's own agents too.
We earn money from AgentStripes through sandbox testing, where we run the agent against fake systems and attacks the developer hasn't seen. The developer pays for the test whatever it finds. A test can lower a rating as easily as raise it, and it can't add a safeguard that isn't there.
AI agent governance you can check
Governance policies say what an agent should do. We think an agent that takes real actions should also come with evidence of what it actually does, and anyone should be able to check that evidence.
The evidence weights in AgentStripes are starting values, and we'll measure them again each quarter from real checks. The full method is in our paper on arXiv. We'd like people to read it and tell us where it's wrong.
Check your agent. Paste a public GitHub link at agentstripes.com. It takes about half a minute and needs no account. Agents with 2 stripes or more can be listed on Agent Hub.