Why AI Coding Costs Explode: The Engineering Behaviours Behind Token Waste

Retry without diagnostic progress

A retry loop becomes expensive when the system keeps attempting the same class of fix without receiving materially new diagnostic information. That pattern often shows up as repeated prompts, the same failing test, and more repository exploration without a changed hypothesis.

TraceYield treats that pattern as a behaviour to inspect, not a blame signal. The useful question is whether the trajectory changed because new evidence entered the loop.

Context waste is usually a task-locality problem

Relevant context can be small. Unhelpful context can be large. The wrong assumption is that more files automatically mean better work.

A narrow task can need four files and 12,000 tokens. The same task can also trigger broad exploration, repeated context loading, and 168,000 tokens. The difference is not just scale. It is whether the agent stayed close to the task.

Requirement quality changes the entire trajectory

A vague instruction like Make the payment endpoint secure invites repository exploration, assumption making, corrective prompts, and extra token consumption.

A specific instruction that names the route, the signature header, the secret, the duplicate-processing rule, and the response contract gives the model a smaller search space and a better chance of a first-pass useful result.

Model selection matters, but not as a price label

An expensive model can be the correct choice for architecture work, ambiguous failures, or difficult debugging. A cheaper model can also be the correct choice for repetitive tasks and constrained edits.

The relevant question is not whether a frontier model was used. It is whether the model matched the work and whether the trajectory justified the cost.

Why normal usage dashboards are not enough

Usage analytics can tell you tokens, spend, models, and users. Engineering analysis needs more: why context was loaded, why retries were necessary, whether diagnostic information improved, whether requirements were sufficiently constrained, and whether the same repository areas were explored repeatedly.

TraceYield is being built to make that difference visible. Instead of stopping at raw consumption, it evaluates the trajectory behind the number.

From AI usage to engineering evidence

TraceYield analyses AI-assisted software engineering trajectories and groups evidence from organisation down to feature-level behaviour.

The goal is not to punish developers for using AI. The goal is to explain why AI usage differs and what engineering behaviour can be improved.

References and method