Temporal Policy

A temporal policy is a rule that has an AI agent remember everything it has done so far in a session, then decides whether to allow its next action based on that whole flow. It's similar to how a store security guard can't spot something suspicious by watching just one moment of a customer's behavior — they need to see the customer's entire path through the store from the moment they walked in.

The traditional approach judged each request on its own, in isolation. Checking an account balance is fine, sending money is fine, making a payment under the approval limit is fine — so each one passes every time. The problem is that when these individually normal actions are chained together, they can form a dangerous pattern: transferring money to another account right after checking a balance, or splitting a payment into multiple smaller transactions to sneak past a limit and exceed the total cap. A temporal policy catches these patterns by checking what happened in previous calls, how much has been spent so far, and whether the sequence makes sense.

For example, imagine an agent that keeps retrying a failed request all night, burning through cost with each attempt. Each individual attempt might look perfectly fine on its own, but a temporal policy that watches the accumulated flow can step in and block it.

The article explains that "a temporal policy looks at the actions an agent has already taken during a session before deciding whether to allow its next call." A common point of confusion: this isn't a generic AI safety filter — it's a specific feature added to Amazon Bedrock AgentCore. The feature is powered by Dogwood, an open-source policy language.