Developer security startup Arcjet Labs Inc. today launched agent runtime security, a new offering that tracks what artificial intelligence agents do once they are running in production systems.

Security teams get a list of which agents are active, and each action an agent takes can be checked against policy before it goes through. Every run is also logged for later review.

The release takes aim at a problem that has grown as agents move beyond chat interfaces. In production they can now write to databases, answer support tickets and process customer refunds, and many of those workflows start from an email or a code commit before moving across several systems with nobody checking each step.

Founder and Chief Executive David Mytton said a risky outcome can build up over “a series of steps that look perfectly reasonable on their own.” Arcjet links those steps and gives teams policy controls that can act at what he called “machine speed.”

Getting agent activity into Arcjet does not require code changes or another agent on the host. Platform teams can send it through the OpenTelemetry tooling they already run, and Claude customers have a second route through Anthropic PBC’s Compliance API. Each agent then appears in an inventory. Prompts, tool-call parameters and security decisions from separate sessions are stitched into one workflow per agent, so an investigator can step through a run one action at a time.

Security teams can set rules to detect prompt injection or redact personal information before it leaks. Bot detection and rate limits are available too. Those policies run on Open Policy Agent and can be changed without redeploying the application, so a team could cap the dollar value of any refund an agent issues without waiting on a release.

Each policy check happens before the action executes. Arcjet returns a decision, and the application can then halt the call, ask a person to approve it or send the agent an explanation. A second check after the call verifies the result before the workflow moves on. The company lists native integrations with the Claude Agent SDK, the OpenAI Agents SDK, LangChain, Mastra and Microsoft Corp.’s Agent Framework, among others.

For audits, correlated traces keep the actions, inputs and policy evaluations from each execution. Some controls, personal data detection included, run entirely inside the customer’s own process, so that data never leaves its environment. Customers can keep audit records in Arcjet’s cloud, in a single-tenant or private cloud deployment or in storage they manage themselves.

The release also moves Arcjet’s Python and Go software development kits to version 1.0. The JavaScript and TypeScript SDK got there in February and now supports more than 15 frameworks, including Next.js. A separate guard interface that is not tied to any framework extends coverage to background jobs and other code, including processes that never handle a web request.

Arcjet says more than 500 production applications now run its software, three years after Mytton founded the company. Its backers include Plural, which led an $8.3 million Series A round last October, as well as Andreessen Horowitz, Seedcamp and former Twilio Inc. Chief Technology Officer Ott Kaukver.

Image: Arcjet