AI Operations · Glossary
Backtracking
Backtracking is the authored capability of an Interactive Agent to revisit a prior routine step when newly available context changes the right answer, governed by the policies in scope at the step the agent returns to.
Linear traversal assumes the right step is the next step. Backtracking accepts that the right step is sometimes the prior step, once new context has arrived.
The capability is authored. The domain expert marks steps where backtracking is permitted, defines the conditions under which it is triggered, and names the policies that govern the revisited step. The agent does not invent the capability. The author configures it.
Backtracking is not retry. Retry repeats the same step with the same context. Backtracking returns to a prior step with materially new context and re-evaluates the path forward.
The trace records the backtrack as a first-class event (see Trace). Per-decision traceability extends to the revisit. An evaluator can later inspect why the agent revisited and whether the policies governing the revisit held.
Backtracking is one of the operating-model affordances that distinguishes Routine Adaptive Traversal from rigid sequence execution.