Research / Coverage area 01
Coverage area 01AI Coding & Development Agents
Coverage of the tooling that lets a five-person engineering team hold a roadmap that used to need twenty: pair-programming assistants, code generation, review automation, and agents that carry a ticket end to end. We track what is actually running in production, what it displaced, and where the throughput gain stops.
At what point does adding an AI coding agent stop adding output — and what did the teams past that point change about how they work?
Sample Data
Three figures from this area's dataset, in the form they publish in.
Illustrative sample — these figures show the structure of the dataset and the form a published finding takes. They are not published findings. Sourced figures publish with the report, each carrying an openable source.
The third figure is the one that matters. Adoption is near-universal and velocity claims are loud; almost nobody is measuring what the change did to quality. Any velocity number published without a paired defect measurement is an unverified claim, and we score it that way.
What We Track
- Deployed tool and model per company, with detection date and source
- Engineers per shipped release, and the trend across four quarters
- Roles not backfilled after adoption — the displacement signal
- Seat cost per engineer per month, list vs. actual where disclosed
- Whether defect rate or change-failure rate is measured at all
In scope
- Assistants and agents used against a production codebase
- Code review, test generation, and migration automation
- Agent orchestration inside the SDLC (ticket → PR → merge)
- Pricing, seat economics, and the real per-engineer cost
Not covered
- Consumer or hobbyist coding tools with no team deployment
- Model benchmarks — we cover deployed outcomes, not leaderboards
Adoption Depth Rubric — 0 to 5
The Leverage Index scores automation and agent adoption depth at 25% of the composite, applied as a 0–5 rubric to each coverage area and then averaged. This is that rubric for this area. A capability scores above 2 only if it is in production, used at least weekly, and displacing or avoiding a named role — not piloted, not licensed-and-idle. See the full inputs and weighting →
No assistant in the development workflow.
Seats purchased, individual and occasional use, no team practice.
Daily completion and generation use by most engineers. No workflow change.
Agents run in CI or review — test generation, review passes, migrations — on every change.
Agents carry defined ticket classes to a reviewable PR unattended. A named role was not backfilled.
Team shape and roadmap capacity were rebuilt around agent throughput, with quality instrumented.
C.A.S.E. in This Area
The same four-stage method runs in all six coverage areas. Here is what each stage actually does in this one.
Agents watch changelogs, release notes, public repository and package-registry activity, engineering job postings, and vendor pricing pages nightly. A stack detection is stored with the page and timestamp that produced it.
Every vendor velocity claim is checked against the customer's own shipping record. Claims with no disclosed baseline are flagged, not averaged in. Conflicts between a vendor case study and observed release cadence go to an analyst.
The AI Coding Agents landscape and the engineering slice of the Leverage Index: where the throughput curve flattens, and what the teams past the flattening point changed.
Inquiry for engineering leaders on the build-vs-buy call, the seat-economics case, and what the adoption depth rubric says about where you actually are.
Sample Insights
Findings in the form we publish them: each one carries a confidence level, and the ones the evidence does not yet support are labeled open rather than quietly dropped.
Velocity gains concentrate in greenfield and migration work. Teams reporting gains on legacy maintenance are largely reporting time-to-first-draft, not time-to-merged.
The binding constraint moves to review capacity within two quarters of serious adoption. Teams that did not add review automation gave back most of the gain.
No credible public dataset yet links agent-authored code volume to change-failure rate. We are collecting it; we will publish the null result if that is what we find.
Illustrative sample — these figures show the structure of the dataset and the form a published finding takes. They are not published findings. Sourced figures publish with the report, each carrying an openable source. Research integrity policy →
What Publishes Next
Want this area applied to your decision?
Book an inquiry with the analyst covering it, or commission research on a question this area does not answer yet.