agentic loop

In depth

One turn of the loop

A turn of the loop always follows the same order: the harness assembles what the model is to read, the model chooses an action, the harness carries it out through a tool, then the result returns into the context. The next turn sets off from this enriched state, so that the output of one turn becomes the input of the next. Nothing is retained by the model between two calls: continuity is carried entirely by the text the harness hands it. This is why an agentic loop reads as an accumulation, where each turn adds its observations to the context until it saturates.

Knowing when to stop

A loop needs to know how to stop, and that is the part people forget to design. Three exits exist: the goal is judged to have been reached, a limit is crossed (number of turns, budget, duration), or an error blocks the system. Leaving the model as sole judge of whether the goal has been reached is the most common flaw, for it readily declares an incomplete piece of work finished. Serious loops therefore set a criterion that can be verified outside the model: a test that passes, a field filled in, a total that comes out right.

The traps

The first trap is the loop that turns without advancing, when the model repeats the same action, obtains the same error, and rewords without changing anything. The second is silent drift: a false observation enters the context at some turn or other and contaminates all the following ones, with nothing to signal it. The third is cost, for the context grows at every turn and every turn is paid for, so that a long loop comes to more than the apparent sum of its steps. These three traps explain why a loop is designed with traces, hard limits and points of resumption, rather than with trust in the model’s judgement alone.