Plans That Adapt

Military strategists have a saying: no plan survives first contact with the enemy. Scientists say something similar: no theory survives its first experiment unchanged. And software developers know it too: no design survives its first users. What all of these fields have discovered independently is that reality is messier than any plan anticipated — and the ability to adapt to that messiness is what separates successful agents from ones that simply fail gracefully.

Why Plans Go Wrong

Plans break down for a handful of predictable reasons. The environment changes: the flight the agent booked gets cancelled, a webpage the agent was scraping goes offline, a database record the agent expected is missing. An assumption was wrong: the task took twice as long as estimated, a dependency turned out to be non-existent, the output of a previous step was not in the format the next step needed. New information arrives: a better approach becomes apparent midway through, or a constraint that was not mentioned at the start is revealed. None of these failures means the original plan was bad. Plans are necessarily made before full information is available. The problem only arises if the agent treats the plan as sacred — continuing to follow it even when it is clearly no longer appropriate.

The opposite of plan fixation is replanning — generating a new plan, or adjusting the current one, in response to a changed situation. Replanning does not mean starting from scratch every time something goes slightly wrong. It means monitoring the plan's progress against reality and making targeted adjustments when discrepancies appear. A well-designed planning agent does three things continuously: it executes the current plan, it monitors whether the world matches its expectations, and it triggers replanning when the gap between expectation and reality is too large to ignore.

Strategies for Adaptive Planning

The simplest replanning strategy is to detect the failure, discard the broken section of the plan, and generate a new path to the goal from the current state. This is called repair planning — like patching a road rather than rebuilding the whole highway. A more sophisticated strategy is contingency planning — building branches into the plan in advance. Instead of a single linear sequence, the plan includes: if X happens, take this branch; if Y happens, take that one. This is how airline operations work: there is always a backup gate, a backup crew, a backup routing. Contingency plans cost extra to build but save enormous time when the primary path fails. AI agents can also use a strategy called plan monitoring with thresholds: the agent keeps executing the current plan but tracks a set of conditions. If any condition drifts beyond a set tolerance — say, estimated cost exceeds budget by 20% — it pauses and replans only that affected section.

The Cost of Replanning

Replanning is not free. Every time an agent stops to recalculate its plan, it spends time and computational effort. If the agent replans after every tiny deviation, it never makes progress — it spends all its energy planning and none executing. This is called thrashing: the agent oscillates between planning and replanning without ever getting anything done. The skill of adaptive planning is knowing when to replan. Minor deviations — a step took slightly longer than expected, a small piece of data was formatted differently — should usually not trigger replanning. Major deviations — the whole approach is blocked, a critical resource is unavailable, the goal itself has changed — must trigger replanning. Well-designed agents use thresholds, priority levels, and explicit monitoring rules to make this judgment automatically. Just as a ship's autopilot makes small constant corrections without the captain's intervention, but hands control back for a hurricane, a good agent handles small surprises silently and escalates big ones.