Coding got 8.3 times faster. The team got 65% faster.
One backend developer on a long-running client platform got 8.3 times faster at coding over eighteen months of AI tools. The team around him delivered 65% more through the same path. That gap is the ceiling of AI tools in a conventional team, and this engagement measured it month by month.
A project that started before agentic tools
The client is a multi-office US property valuation firm. First Line Software builds its core platform turnkey, and the work has run for two years. When the project started, AI in an editor meant code completion. Agentic tools did not exist yet.
That makes the project a rare instrument. The same product, the same codebase and the same estimation scale span the whole arrival of AI in software delivery. The comparison is an engineer against his own earlier self.
Hours per story point, window against window
The series runs from April 2025 to September 2026, one point a month, for one backend developer. Hours come from git, Azure DevOps and timesheets. Work is sized in story points on the scale the project used from its first stages, the original story points. The measure is hours divided by story points.
The baseline window, April to September 2025, averages 4.49 hours per story point. The window after Copilot settled, January and February 2026, averages 2.36. The last window, July to September 2026, averages 0.54.
That is 8.3 times, window against window. It is not the first month against the last, which would give a larger and less honest figure.
The series carries its own caveats. May 2025 sits off the scale because of leave, and keeping it in is the conservative choice: without it the gain is 7.8 times. November 2025 is an estimate, not a measurement, because a one-off data migration was taken out of it. September 2026 runs only to the 22nd.
Each tool cost a quarter
There were two tool switches. Copilot arrived in October 2025, and the gain landed in January 2026. Claude Code came into use in April 2026, and the gain landed in July. Each switch took about a quarter.
Adoption was not automatic. Engineers resisted the tools at first, and it took time to make them part of the expected skill set. A tool that changes what an engineer has to know costs a quarter before it pays.
The rest of the path barely moved
The core platform team counts ten people. Across the same period, the whole delivery path gained 65%, measured by emulating that path inside the Data Extractor stream. Coding fell 8.3 times.
Requirements, analysis and architecture got a little faster. So did testing, stabilization and release. The team used AI in those stages too, but as a supporting tool, and the gain there was small.
So 8.3 times in one segment became 1.65 times end to end. In the terms of the RACE triangle, the team took the gain as scope: 65% more finished work through the same path. That is a real result, and it is also where tools stop.
The constraint is the hand-off
The arithmetic explains the ceiling. Once coding takes a sliver of the path, the remaining time sits in the stages around it and in the hand-offs between them. Every task still changes hands between the people who own those stages. A faster developer does not shorten a queue that forms before and after him.
Everyone talks about a gain like this one. This engagement measured it, on one project that ran from before AI tools to agentic tools in an engineer's hands.
The team, not the tools
The 65% decided the next step. From September 2026 the same project runs with a different team structure: a micro-team of an AI Product, a Forward Deployed Engineer and an AI-fluent Quality Engineer.
The AI Product talks to the client and turns intent into specification. The Forward Deployed Engineer designs the system and builds the machinery the work runs on. The Quality Engineer owns the gate rather than acting as a station the work passes through. Both the AI Product and the Quality Engineer work with AI teammates First Line built in house.
Requirements, coding, testing and release now sit with one micro-team, and nothing is handed outside it. That is the RACE Programming answer to the ceiling: change who carries the work, so there are fewer boundaries for it to wait at.
There is no data from the new structure yet. The restructured team is the experiment, and the next measurement will say whether it moves the 65%.
Client identity withheld. The same firm runs a second stream that was built in the new shape from its first day: agent-to-agent data imports, by two people and the agent. A third stream took AI onto the firm's own desks: an agentic workspace for every appraiser.