Goal Persistence as Inertia: When an Agent Should Abandon a Failing Plan

A coding agent chooses an implementation approach, runs the tests, and gets a failure. It patches the same approach. The next run fails differently. It patches again, adds another workaround, and consumes several more tool calls without reconsidering the underlying design. By the time it finally replans, much of the remaining context and execution budget has been spent defending a path that stopped being plausible several steps earlier.

The reverse failure is just as recognizable. An agent selects a reasonable approach, encounters one failed test or one ambiguous tool response, abandons the plan immediately, and starts again from a different direction. The new plan is no better; the agent simply has too little resistance to changing course.

These cases can look like opposite forms of model unreliability. One appears stubborn and the other erratic. Architecturally, however, they are variations of the same problem: commitment to the current plan is poorly calibrated to the evidence available.

The relevant question is therefore larger than whether the agent can generate a better next action. It is how much evidence should be required to preserve or overturn the plan that organizes those actions. Unless that decision is represented explicitly somewhere in the architecture, persistence emerges from an opaque mixture of prompting, context history, model behavior, memory design, and whatever replanning logic happens to surround the model.

Part 1 used Jon Doyle's Extending Mechanics to Minds as an intellectual catalyst for mechanics-style reasoning about behavior. The framework itself is this series' architectural synthesis. Here, the useful concept is inertia: resistance to a change in state.

Applied to agent architecture, plan commitment can be treated as resistance to abandoning the current course of action. That framing turns persistence from a personality-like description of the model into a design property that can be represented, observed, and governed.

That resistance is necessary. An agent with no persistence would replan whenever a local observation became inconvenient. Multi-step work would be difficult because each failure, delay, ambiguous result, or partial contradiction could erase the continuity required to finish the task. Research would restart after every weak source. Code generation would switch architectures after every test failure. Tool-driven workflows would repeatedly discard useful intermediate progress.

Too much resistance produces the opposite problem. The plan survives evidence that should have invalidated it. New failures are interpreted as reasons for another patch rather than reasons to reconsider the approach. Prior work becomes an argument for continuing simply because it already exists. The system remains organized around a decision whose evidentiary basis has eroded.

Inertia is therefore neither a desirable nor an undesirable property on its own. The design issue is calibration. A useful agent needs enough commitment to carry coherent work through ordinary friction and enough responsiveness to change direction when the accumulated evidence warrants it.

An explicit commitment mechanism needs to account for at least three things: evidence supporting the current plan, signals that the plan is failing, and the prospective cost of switching away from it. This is an architectural model proposed here, rather than a formulation attributed to Doyle.

Confidence and commitment should also remain distinct. An agent can have only moderate confidence in a plan yet rationally remain committed because the alternatives are weaker or changing course would be costly. Conversely, high confidence should not make commitment irreversible when new evidence attacks a foundational assumption. Confidence describes the evidentiary case; commitment governs whether the current course survives into the next step.

Evidence Supporting the Current Plan

Commitment should strengthen when observations continue to support the assumptions on which the plan depends. Tests pass. Tool results confirm expected conditions. Intermediate artifacts remain consistent with the intended design. Dependencies behave as anticipated. Progress toward the goal remains visible.

The important distinction is between evidence for the plan and effort already invested in it. Ten successful intermediate checks can legitimately increase confidence that a plan remains viable. Ten expensive steps merely spent pursuing it do not provide equivalent evidence. If an architecture allows sunk effort to masquerade as confirmation, persistence can increase exactly when the case for the plan is weakening.

Failure Signals Against the Plan

Commitment should weaken when failures accumulate, but failure counting alone is too crude. A transient timeout, a deterministic test contradiction, and repeated evidence that a core assumption is false should not exert the same pressure.

The useful architectural question is what each observation says about the plan itself. Some failures indicate execution noise while leaving the underlying approach intact. Others attack a dependency or assumption on which the plan cannot succeed. Replanning logic becomes more reliable when the system can distinguish local recoverable errors from evidence that the current strategy is structurally wrong.

This also means that repetition matters. A single weak signal may justify another attempt. Several independent failures pointing toward the same invalid assumption should exert progressively greater pressure against continued commitment. Without some form of accumulation, an agent can repeatedly treat the latest failure as isolated and never register that the overall evidentiary position has changed.

The Cost of Changing Course

Replanning is not free. The relevant quantity is the prospective cost of changing course from the current state: useful work that would be discarded, new tool calls, reopened decisions, additional context consumption, or inconsistency with actions already taken in external systems. A plan that is only slightly worse than an alternative may still be worth completing when switching now would impose substantial cost.

That is different from sunk cost. Effort already consumed is not evidence that the plan remains viable and should not strengthen commitment simply because it was expensive to reach the current state. Prospective switching cost can legitimately raise the threshold for change; sunk effort cannot repair a deteriorating evidentiary case. The architecture needs both questions kept separate: how credible is the current course, and what will changing it cost from here?

Taken together, these three inputs provide a practical model of commitment. Supporting evidence increases resistance to change. Failure signals reduce it. Switching cost raises the amount of contrary evidence required before changing course makes sense. Once these variables are made explicit, persistence stops being an accidental behavioral residue and becomes something the system can observe and govern.

When commitment is implicit, debugging tends to stop at descriptions such as “the model got stuck,” “the agent was stubborn,” or “it replanned too aggressively.” Those labels describe visible behavior without identifying the architectural variable that produced it.

A state-based diagnosis asks a different set of questions. What evidence was still supporting the plan when the agent persisted? Which failure signals had accumulated? Were those failures represented in a way the replanning mechanism could use? What would switching have cost at that point? Which part of the architecture determined that the existing balance still favored persistence?

This matters because improving the next prompt may change behavior without making the mechanism clearer. A prompt that says “reconsider your approach after repeated failures” leaves several consequential decisions unresolved. What constitutes repetition? Are all failures equivalent? Does a contradiction in a foundational assumption count more than a failed execution attempt? Should a nearly completed workflow tolerate more friction than a newly selected plan? When should irreversible external actions make the system more cautious about switching?

Those are architectural policy choices. They can be represented in different ways: explicit state, orchestration logic, plan metadata, evaluation gates, memory structures, or other control mechanisms. The implementation can vary. What matters is that the commitment decision has a defined home and observable inputs.

Once it does, persistence failures become easier to classify. Excessive commitment means contrary evidence is not reducing commitment quickly enough, or switching cost is being overweighted. Premature abandonment means supporting evidence is being forgotten, local failures are being overweighted, or the cost of resetting the plan is not represented. The debugging target becomes the commitment mechanism rather than a vague judgment about model temperament.

A single persistence profile is unlikely to be appropriate across agent workloads. Different task classes impose different costs for persevering and different costs for switching.

A customer-facing agent operating in a short interaction may need relatively low commitment to an uncertain interpretation because persisting with a mistaken assumption can quickly propagate into an incorrect answer or action. Reconsideration can be cheap because little state has accumulated and the cost of revisiting intent is modest.

A long-running research agent has a different shape. It may have gathered evidence across many sources, created intermediate analyses, and established a working hypothesis that should survive isolated contradictions. Replanning the entire investigation after every weak signal would destroy continuity. At the same time, repeated independent evidence against the central hypothesis should eventually override the value of preserving prior work.

A coding agent may require still another profile. Early architectural decisions can be cheap to reverse, while later changes may invalidate many dependent edits. But increasing switching cost should not make the agent indefinitely defend an implementation once tests repeatedly demonstrate that a foundational assumption is wrong.

The practical implication is that commitment should be visible and tunable as part of the agent's operating design. Teams should be able to identify what strengthens commitment, what weakens it, how evidence accumulates, what switching cost means for that task class, and what threshold causes the system to reconsider its plan.

That threshold need not be one universal number or formula. It may be implemented through rules, evaluators, state machines, confidence structures, or orchestration policy. The architectural requirement is simpler: persistence should result from an intentional relationship between evidence and the cost of change. Once that relationship is explicit, the system can be tested for over-persistence and premature abandonment as distinct failure modes rather than discovered unpredictably in production behavior.

This is part 2 of 4 in the series The Mechanics of Agentic Behavior.

• Part 1: Why Agentic Systems Need a Theory of State Beyond the Reward Function
• Part 3: Competing Objectives as Forces: Designing Agents Under Multiple Pressures
• Part 4: Correct Actions, Unstable Trajectories: Reasoning About Agent Behavior Over Time

Browse the full Architectural Insights series →

• Jon Doyle, Extending Mechanics to Minds: The Mechanical Foundations of Psychology and Economics, Cambridge University Press, 2006.