Composio Firewall screens every agent action with Jev

Karan Vaidya introduced Composio Firewall, a runtime boundary that evaluates every action an agent takes using Jev plus small language models Composio tuned on its own real-world usage data. It checks in both directions, the actions agents take and the information coming back to them through Composio. Vaidya says it is live now for all Composio developers.

Karan Vaidya

@KaranVaidya6

X

every action an agent takes is evaluated using Jev and custom SLMs tuned by Composio based on real-world usage data

Karan Vaidya

@KaranVaidya6

X

In traditional software, those were set at build time. With agents, they have to be set and enforced at runtime.

The pitch is that build-time boundaries do not survive contact with agents. Composio already sits in the path as the runtime that hands agents tools, auth and sandboxed execution, so the firewall is an extension of a boundary that was already there rather than a new hop. Vaidya frames the job as verifying that the agent respects user intent and the boundaries set by the developer and user, with protection against prompt injection and other attacks listed as the safety payoff.

What is not in the announcement

No numbers. There is no latency figure for the added evaluation, no cost per screened action, no catch rate against injection attempts, and no false positive rate for legitimate actions that get blocked. There is also no detail on how work is split between Jev and the custom SLMs, which matters if you are trying to reason about the per-call price of the thing. Everything here is a vendor description of a vendor product, not a measured result.

For people building on it, the question to ask before wiring this into a hot path is what it does to a multi-step agent run. An agent that makes thirty tool calls in a session pays the firewall thirty times on the way out and again on the way back, and that compound cost is exactly what the post does not quantify. A guard that is cheap on a single call can be the dominant line item on a long trajectory.

This is the second product we have covered that puts Jev in front of tool calls as a guard, after OpenRouter's Armor ran on Jev 1.13 earlier. The pattern is becoming the default answer to agent safety, and so far nobody shipping it has published what the screening layer costs or how often it is right.