Runtime Control for Tool-Using Agents
Our preliminary DARC study compares ways to review actions using converted examples from four public agent-safety benchmarks. The report explains the evaluation, results and limitations.
Read the research noteAI that acts.People in control.
We study how AI agents behave and build tools that help people understand and review their actions.
Meet Ordin.
Understand what a command will do before you or your agent runs it. Ordin reviews actions locally and returns a decision with reasons you can inspect.
See the action.
Understand the impact.
OrdinAction review
Proposed action
Show me what changed in this project.
git status --shortEffectReads repository state
ContextCurrent project
Allow
No files are changed by this command.
Reasons you can inspect
Illustrative review. Your policy and context determine the decision.
See what a command will do.
An agent wants to check a project. A review can identify a read-only command and explain why it can continue.
See how to integrate itThe research
behind the review.
We study how tool use, permissions and action history affect an agent's decisions. The goal is useful evidence about when an action should proceed and when a person needs to step in.
These are preliminary results on converted benchmark examples, not proof of safety in live deployments. Explore our research
Find your
starting point.
Ordin
An open-source Python library and command-line tool for developers and AI agents. Core reviews run locally without sending command data to a hosted service. Your runtime stays in charge of execution.
Get started on GitHubRuntime Control
Our platform preview gives teams a shared workspace for reviewing AI actions. Apply policies, route sensitive actions for human approval, and keep records of what was decided and why.
Explore the platformStart with
the basics.
What is an AI agent?
An AI agent is software that can use tools to carry out a task, such as editing a file, running a command or sending a message. Those actions can change things outside the conversation, which is why reviewing them matters.
How do LatentOps and Ordin fit together?
LatentOps is the AI safety research and product company. Ordin is our open-source project for local command and tool-call review. Runtime Control is our platform preview for teams that need shared policies, approvals and records.
Does Ordin run my commands?
The review library does not execute commands. Some integrations can run or forward an action after the review permits it. Your application remains responsible for execution, approval flows, credentials and sandboxing.
Does a review guarantee that an action is safe?
No. Reviews depend on recognized behavior, your policies and the context you provide. Keep permissions, sandboxing and human checks in place. Our research reports its test conditions and limitations rather than treating benchmark results as proof of deployment safety.
Where should I start?
Watch the Ordin demo to understand the idea, then follow the GitHub quickstart if you want to integrate it. If your team needs a shared review workspace, explore the Runtime Control preview or get in touch.
Put a review
before the next step.
Start with Ordin, or talk with us about the agent workflows you're building.