temporal policies

Temporal policies are rules that decide whether to allow an AI agent to do something right now by also checking what it did before, not just looking at the current action in isolation.

Here's an analogy. One security guard just checks the ID of whoever shows up at the door and lets them in. Another guard remembers what that person did inside a few minutes earlier, and stops them if what they're trying to do now doesn't match up with that. Temporal policies work like the second guard. An AI agent decides for itself and carries out a sequence of steps, and each step can look fine on its own, yet look suspicious once you connect it to the steps before it. For example, the account number the agent looked up might not match the account number it's now trying to transfer money to, or it might approve an insurance claim seconds after rejecting that same claim.

These policies run outside the program code that drives the agent, at the gateway where requests actually pass through. So even if there's a bug in the agent's code, or someone tries to trick the agent with a clever prompt, the gateway itself is hard to bypass.

The article reports that AWS added temporal policies to Amazon Bedrock AgentCore. Existing policies checked each request in isolation, but the new feature is different in that it also looks at the sequence of prior tool calls. A common misconception is that this is a safeguard built into the agent's code, but it actually runs at a gateway outside the agent's code, filtering requests regardless of code bugs or manipulation attempts.