When AI Agents Should Repair, Replan, or Abandon: A Threshold Theory of Plan Commitment
Organizations: Omega Venture Partners www.omegavp.com
Abstract
AI agents often discover mid-execution that their plan is broken, for example when an action fails or a tool is withdrawn. Deployed systems use conventions such as retry loops or "always replan," and a recent benchmark finds no fixed strategy dominates. This paper derives the optimal policy. In the model, a disruption destroys the causal support of part of a plan; repair is a partial reset that preserves the plan's structure, replanning is a full reset, and changing a plan is costly because others have conditioned on it. The optimal policy is a threshold in effective disruption depth: repair shallow damage, replan after deep damage, and abandon the goal when even a fresh plan is not worth its cost. Depth depends on plan structure. In expectation, a single broken action forces 1+R(G)/n actions to be revalidated, where R(G) counts the pairs of actions joined by a dependency path; to first order, a fully serial plan switches to replanning at raw damage about w times smaller than a plan of w parallel chains. A closed-form stability-price threshold determines when a worse repaired plan should be kept over a better fresh one. As language-model planners push replanning costs toward zero, the stability price becomes the only force sustaining plan commitment. Under repeated disruptions, revision frictions make commitment hysteresis optimal: the agent keeps executing a plan it would never choose fresh, and each repair adds patch debt until the debt outweighs repair's cost advantage and the agent rebuilds. A value-of-diagnosis result states when measuring damage is worth its compute and shows that "always replan" is optimal when damage is predictably deep and replanning is the better fallback. A separately implemented numerical checker tests each proposition (zero violations; 1.6 million static-model evaluations), and the model's regime map reproduces that benchmark's regime flips qualitatively.
Figures & tables
| Symbol | Meaning | Source in deployment |
|---|---|---|
| , | remaining plan size; dependency DAG | plan representation |
| , | broken actions; raw damage | failed precondition checks |
| , | validation footprint ( 1 ); effective depth | reachability over |
| damage amplification ( 6 ) | computable at plan time | |
| , | post-disruption fresh optimum; outside option | assigned |
| , | repair / replanning compute cost | token and latency logs |
| Component | Claims tested | Scale | Outcome |
|---|---|---|---|
| Symbolic audit | Props. 1 and 3 closed forms | 8 checks | all pass |
| Static scan | monotonicity, uniqueness, regimes, -statics | 3,946 inst. 401 grid | 0 violations |
| Scale corollary | Corollary 1 | 395 (294 admissible) | 0 violations |
| Revision risk | App. B single crossing | 3,959 instances | 0 violations |
| Graph enumeration | ( 6 ), bounds, extremes, parallel | 600 random DAGs | exact ( ) |
| Sequential DP | threshold structure; patch-then-rebuild | 59 draws ( grid) | 59/59; 26/26 |
| Reported finding (authors’ own) | Model counterpart |
|---|---|
| Repair changed of the original first-step actions; replanning (DriverLog-Time) [ 10 ] | the premise |
| Repair’s advantage grows with problem size [ 10 ] | Corollary 1 (scale widens the repair region) |
| Repair preserved – of itinerary content vs – for replanning [ 40 ] | the premise, second domain |
| Replanning beat hierarchical repair only at compound disruptions with the strongest replanner [ 40 ] | regime flip at high , low effective replanning burden (Figure 6 ) |
| Success-rate tie at single disruptions, strongest replanner, with repair more stable [ 40 ] | cell uninformative on success alone; decided by the stability term ( ) |
| No fixed strategy dominates [ 40 ] | parameter-dependent regime taxonomy (Proposition 1 ) |