Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

Why Does LangChain's New Jev Auto Mode Forget Your Request at Tool Call 16?

​

After introducing Jev and the idea of using it as a guardrail classifier in LangChain’s Auto Mode, the author explains how they reproduced the problem without a TypeSafe key by pointing the middleware at a local stub and capturing the classifier payloads. They show three concrete issues: (1) the middleware builds Jev’s “authorization state” from only the last 30 messages, so the explicit user message drops out mid-run (around tool call 16 in their setup), leaving the guard to judge with no user text present; (2) LangChain’s default risk criteria are never actually sent because the middleware constructor defaults its criteria argument to None and overwrites the config’s defaults, so Jev receives only the “risk criteria sent” flag; and (3) rate-limit or service errors (429/503) cause the agent run to fail immediately (“fail-closed”) with no retry. The author then provides a concise fix: pin the newest user message into Jev’s classification state for the full run length and explicitly pass the desired Noul criteria, plus guidance on retrying around rate limits outside the middleware. Finally, they recommend using Auto Mode only for short tasks and using the pinned variant (with human-in-the-loop for destructive tools) for longer agent runs, concluding that the integration’s experimental one-liner currently ships an authorization check that can lose the user and a risk definition that may not leave your machine.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI