Steve Jamison studies how coding agents change the economics and practice of software development. His work at Jellyfish connects engineering data with hands-on agent workflows: how to compare rising AI bills with developer output, encourage useful adoption and give agents the context needed to review their own changes.
From biomechanics to engineering data
Steve trained in mechanical engineering, earning a Ph.D. at The Ohio State University. Before moving into startups, he researched injury mechanisms and rehabilitation at the Center for the Intrepid and Harvard Medical School. His early research examined how trunk movement affects forces on the knee, with implications for understanding anterior cruciate ligament injuries.
He was the first author of a 2012 randomized study of trunk stabilization exercises that tested whether adding those exercises to resistance training improved strength, movement control and knee loading. The added exercises did not improve several targeted outcomes relative to resistance training alone. Steve and his collaborators suggested that training involving unexpected disturbances and dynamic movement might better address the problem. The distinction mattered: improvement in an exercise measure did not necessarily mean improvement in the movement the intervention was intended to change.
Steve’s engineering cost analysis examines why a growing AI invoice can coexist with a falling allocated cost per merged pull request. Using Jellyfish’s engineering data, he reports a 60-fold increase in AI spending, 43% more merged pull requests per engineer and a 26% decline in combined cost per merged pull request over the period analyzed.
The mechanism is a change in the denominator: when salary remains the largest expense, additional output can spread that salary allocation across enough work to outweigh higher model costs. That calculation describes allocated cost per unit of output; it does not establish equivalent cash savings or prove that AI spending caused the throughput increase. Pull-request counts also cannot establish customer value. His argument gives managers a concrete reason to consider total engineering cost, output and quality together before capping token spending.
Agent adoption as behavior change
Steve’s research on coding-agent adoption includes analysis of approximately 780,000 Cursor developer-weeks. He connects the study of engineering outcomes to practical recommendations: share proven skills, explore applications and sustain use. This treats adoption as a change in working habits that teams can help develop, rather than a decision completed when they purchase a tool.
The analysis is observational: patterns across developer-weeks do not establish that agent use caused a particular improvement. His recommendations nevertheless give teams concrete practices to try and assess against their own results. Sharing a useful skill makes a working approach available to colleagues; exploring applications gives developers opportunities to find where agents help; sustained use lets those experiments extend beyond an initial trial.
Bringing the task back into review
Steve’s public account of his engineering workflow describes skills for preparing issues, opening draft pull requests and reviewing changes. His review skill combines the code with repository instructions, the human-written issue and the pull-request description, then produces a prioritized set of edits. It addresses a specific failure he encounters with agents: implementation can miss or lose relevant context, so review needs to restore the original task as well as inspect the resulting code.
He also uses agents to capture decisions and meeting commitments and assemble daily priorities from his work systems. Earlier note-taking routines had failed because recording the work became a separate chore. These experiments test whether agents can make that capture part of the working day rather than another task to maintain.
Steve Jamison uses Jellyfish’s engineering data to explain how a much larger AI bill can coexist with a lower allocated cost per merged pull request—and why throughput, quality and work size belong beside that calculation.
Combined salary-plus-AI cost per merged PR can fall while the AI bill rises, because additional output spreads the much larger salary expense across more merged changes.
The heaviest-spending fifth reported 67% throughput growth and a 32% decline in cost per PR, versus around 40% and 25% for lower-spending peers. These comparisons do not isolate AI’s causal contribution.
Reverts showed no worrying increase, but did not demonstrate quality improvement. Rising lines per PR also complicate the result: more merged changes can bring more review work.
Before adding token limits or investing in cost optimization, compare the likely effect with throughput enablement. Track total cost per merged PR, spending, throughput and quality together each month.
