ML//agent//active perception

Active perception is acting in order to choose the next observation: taking actions that do not bring the agent closer to its goal but reduce its uncertainty about where it is or what is true. A drone detours to look around a corner before committing to a corridor, a warehouse robot stops to relocalize against a known shelf, a radar spends one sweep confirming an echo before assigning it a track. A programmer does not read an entire repository before changing one function either: they search, inspect the likely file, follow a dependency, run a test, and use the failure to decide what to inspect next.

Active perception is acting in order to choose the next observation: taking actions that do not bring the agent closer to its goal but reduce its uncertainty about where it is or what is true. A drone detours to look around a corner before committing to a corridor, a warehouse robot stops to relocalize against a known shelf, a radar spends one sweep confirming an echo before assigning it a track. A programmer does not read an entire repository before changing one function either: they search, inspect the likely file, follow a dependency, run a test, and use the failure to decide what to inspect next.

These are epistemic actions. They reduce uncertainty without completing the task, while moving, editing or deploying are world actions; running a test sits between both. The choice between them is value of information made sequential: an observation is worth taking now when what it may change in the decisions that follow outweighs what it costs in time, energy or risk.

A POMDP is the frame that sees these actions. Deciding on the belief instead of on a single estimate, it can prefer a detour that sharpens the belief, which a planner working with certainty equivalence never does: acting as if the estimate were true, it sees no reason to look. In practice robots get much of the benefit with simple rules, such as relocalizing when the estimate's covariance grows past a bound.

For a software agent the real project state is hidden, tools produce observations, and the policy chooses the next acquisition or intervention. The views it can request are the ones the harness provides (its epistemic interface); active perception is the choice, step by step, of which one to ask for next.

Dumping the whole repository into the context window removes this useful control loop and can bury the evidence that matters, as flying a fixed survey pattern spends battery on places that change no decision.