Guardrails in the Agent Loop: Don't Put Them All at the End

Frank Coyle's Day 1 talk at the AI Engineer Code Summit had a very concrete topic: where guardrails should sit in the agent loop.

Many teams make guardrails an outer filter, reviewing only after the model has finished generating. This placement is useful, but it misses the busiest stretch of the agent's path. The model decides to call a tool, the tool returns a result, and the model continues acting based on that result. Errors often appear between these steps.

The checkpoints presented in the talk can be broken into three layers: validate input before calling a tool, check output after receiving the tool result, and judge whether the result is reasonable before moving to the next step. The three layers sound like common sense, yet they are the easiest to skip in engineering. Input validation can block parameter errors and unauthorized calls; output checks can catch bad data returned by tools; and reasonableness checks can prevent the model from treating irrelevant information as evidence and continuing to reason.

Guardrail placement matters more than quantity. If checkpoints are in the wrong place, adding more rules later only adds latency. Doing validation in the middle of the loop increases the overhead of each tool call and makes agent behavior harder to debug, so it needs to be risk-tiered.

This talk did not offer a one-click solution. What it offered was a sense of placement: guardrails belong between steps, not at the final gate. The smoother the tool calls, the more you need to pay attention to that middle stretch of the path that no one is watching.

Source: AI Engineer tweet