What happened to team throughput as AI agents were adopted, 2025 to 2026

Career·Updated

What happened to team throughput as AI agents were adopted, 2025 to 2026

Matt built the measurement before he made the claim. In February 2026, the same month the team received its first enterprise agent licences, he stood up a metrics pipeline over the issue tracker and the code host so the effect could be observed rather than asserted. This document is what it found, and, just as importantly, what the study itself says about its own limits.

The framing he insists on: this is a correlation story, not a controlled experiment. Everything else that changed in the same window is listed below and should be quoted with the numbers.

Team delivery, code output, and review: the direction of travel

Delivered tickets carrying the team label moved up and down through 2025 without a trend, then rose in every quarter of 2026, the latest quarter being the largest by a wide margin. Merged pull requests for the same five engineers rose in every quarter of the two years and accelerated sharply from 2026 Q2. Reviews given showed no trend across 2025 and then climbed steeply through 2026. Median cycle time to production sat inside a seven- to-ten-day band from 2025 Q1 through 2026 Q2 and then fell sharply in the latest quarter, while average days to merge stayed inside a one-to-five-day band and was at its shortest in that same quarter. The endpoints are in the headline ratios below; the per-quarter detail is not published here.

Headline ratios

| Measure | 2025 full year | 2026 through Sep 12 | Ratio | |---|---:|---:|---:| | Merged pull requests per week, five engineers | 10.3 | 28.2 | 2.7x | | Median cycle time to production | 7.0 to 9.1 days | 2.3 days in Q3 | about 3 to 4x faster |

Over the same window, delivered tickets per week rose about 2.3x and reviews given per quarter rose roughly 2.5 to 3.9x; the absolute rates behind those two ratios are not published here. The pull-request gain holds, smaller, when Matt's own output is excluded.

Quality held while throughput rose

Customer-reported defects created fell steadily while output roughly tripled, taking the open backlog from 28 to 12. Internally found defects held steady quarter over quarter, which is what you want when review volume is rising: the team is finding its own problems at the same rate across far more code. Reverts stayed rare.

Matt's own output

How much an engineering manager still ships is a fair question to ask him: his median cycle time to production was 1.0 day in the latest quarter, measured across 280 tickets, and his merged pull requests went from 16 a quarter to 339, because from June he was running several agents in parallel on decomposition, deployment readiness, and security work.

The cost side, said out loud

Review load more than doubled, and the pooled median time to merge lengthened rather than shrank. More code arrived, review kept up, and merges got slightly slower rather than being rubber-stamped. Automated adversarial review as a repository workflow, added in August, is the mechanism for keeping that true as volume grows.

What can and cannot be measured about usage

The only per-commit AI signal available is one tool's co-author trailer; the other editor leaves no trace, so every usage figure is a floor, not a count. Team commits carrying the trailer went from none in 2025 to a substantial and rising share in 2026, with a dip in the latest quarter. Across the whole organisation, adoption spread from a handful of engineers to most of the organisation across 2026.

Matt's own trailer share understates his usage, because the autonomous remediation and parts of the deployment work committed under other paths. His interactive session activity peaks in May and then falls as work shifted to autonomous runs, which is why he says prompts are not the measure to quote.

Confounders, in the order an interviewer would raise them

  1. Work mix changed. Security and compliance tickets rose sharply, and the deployment programme generated many small tickets. Small tickets inflate counts and shrink cycle time. This is why the story leans on merged pull requests per person and on cycle time rather than ticket counts alone.
  2. Team composition changed over the window, so the team totals are not like-for-like.
  3. Story-point practice changed. Points coverage rose sharply across the window, and the large roll-up epics closed in 2026 are excluded from the points figures, so points are not like-for-like across the window. Points are the weakest signal of the set.
  4. The tracker project changed mid-2026, and ticket-granularity conventions may differ between the two.
  5. Unassigned tickets are numerous and sit in the team totals.
  6. Automation counts as human output. Some agent-opened pull requests sit inside the human counts. Automated dependency bots are excluded by the author filter.
  7. Two different review measures exist, one counting review events and one counting distinct pull requests reviewed. Quote one and say which.
  8. The latest quarter is 74 days, not a full quarter; per-week rates are normalised, totals are not.
  9. Usage telemetry is partial, as above.

What Matt did, as distinct from what happened

He adopted first and in public, using the second editor daily from mid-2025 and giving a lunch-and-learn on it. He ran the tool evaluations honestly, including one that was rejected. He ran the organisation's agent trial in December, secured the first enterprise licences for his group in February, gave up his own seat on the other editor to pay for it, and asked the group to mentor others. He built the metrics pipeline the same month. Then he turned the workflow into shared assets: ticket-drafting skills, the plugin marketplaces, the on-call plugin, adversarial review as a repository workflow, and the autonomous security remediation with a human on every merge. He taught it in three brown bags and the weekly demo series. And he set the rule that agents write and humans review and test, modelling it by carrying a large share of the review load alongside his own output, in the quarter when the team's review volume peaked.

Gaps the study names for itself

A 2025 incident baseline is not available; the defect and revert series are the quality evidence instead. Usage telemetry for the second editor was requested in January 2026 and never became a pipeline. And the obvious next step, running the same queries for other teams that received licences later as a natural comparison group, has not been done.