Why Gaten
A prompt is a request. A checkpoint is a rule.
Once an AI agent can refund, pay, email or change a system, the question isn't what it might say — it's what it's allowed to do. That needs to be decided outside the model, on every action, in rules people can read.
The options
Ways teams try to limit agents today.
Each of these has its place. Here's how they compare for one job: controlling the actions an agent takes.
| Approach | Checked outside the model | Sees the action and its arguments | Human approval built in | Rules business owners can read | Decision audit trail |
|---|---|---|---|---|---|
| Instructions in the prompt“Never refund more than $100.” | |||||
| Content guardrailsFilters what a model reads and writes. | Partly | Partly | |||
| Checks written into each toolif amount > 100: raise … | You build it | You build it | |||
| Framework approval stepsPause the agent and ask a person. | You build it | ||||
| General authorization enginesPolicy-as-code for apps and APIs. | You build it | Partly | |||
| GatenA checkpoint built for agent actions. |
A general comparison of approaches, not of specific products. Capabilities vary by tool and by how it's set up.
Use them together
Guardrails watch what agents say. We decide what they do.
Content guardrails catch harmful or off-topic text. Your agent framework runs the loop. Your identity system says who's who.
Gaten sits at a different point: the moment an agent tries to act. It works alongside the tools you already have, rather than replacing them.
Build or buy
What you'd build to do this yourself.
Checking one threshold in code is easy. Running agent limits across a company is a product of its own.
- 01
A rule language and versioning
Rules people outside engineering can read, with history and one-step rollback.
- 02
Approval routing
Who approves what, how they're notified, timeouts, escalation, and safe defaults when nobody answers.
- 03
Dry-run on real activity
Replaying past agent actions through a draft rule to see what would change.
- 04
A tamper-evident log
Every request and decision, linked to the rule version, searchable and exportable.
- 05
One place for every agent
The same limits and records across teams, frameworks and model providers.
From the blog
Related reading
- SecurityAI agent security: a practical guide for teams shipping agentsAI agent security is about controlling what agents can do, not just what they say. The risks, the controls that work, and where to start.
- SecurityPrompt injection in AI agents: how it works and how to contain itPrompt injection turns AI agents against their owners. How direct and indirect attacks work, and why the best defence sits at the action layer.
- SecurityLLM guardrails: content filters vs. action controlsLLM guardrails filter what models read and write. Here's what they're good at, what they miss once agents take actions, and how to cover the gap.


Compare it on your own agents.
Bring the checks you've written today, and an expert will show what changes.