← ResearchProve · Build log11 min read

Build log 01: a year of building with agents, in numbers

Every month we will publish how our team actually works: what we shipped, what it cost in attention, and where the process broke. This first issue covers January to September 2026 and the first half of the fourth quarter.

We are a small team, and most of our code is now written by AI agents working under our direction. That sentence is easy to say and hard to check, so this series checks it. Once a month we will count what we did across every repository we own, say what changed, and say where it went wrong.

This first issue looks back over the first three quarters of 2026 and then at the fourth quarter up to 17 November, which is close to its midpoint.

How we counted

We mirrored the 80 repositories in our GitHub organisation that had any activity this year and read every commit, pull request and CI run in them. Then we added our issue tracker and the records our agent tooling keeps.

A few rules keep the numbers honest:

  • Commits are counted once, however many branches they sit on. Merge commits, bots and commits by outside contributors to an open-source project we fork are left out.
  • Landed means the commit is on a main branch. Pushed means it reached GitHub on any branch. The gap between the two is work that was squashed, replaced or abandoned.
  • Lines ignore lock files, generated files and data. We also set aside 114 commits that each added more than 5,000 lines, because those are imports and data loads, not writing.
  • Test cases are counted by reading the main branch as it stood at the end of each month.
  • The fourth quarter is 48 days old of 92. Where we give a full-quarter figure for it, it is a straight projection and we say so.

Lines of code and commit counts measure activity, not value. We report them because they are the numbers people ask for, and because how they move tells you something. We do not treat them as a score.

The year in one table

Q1 Q2 Q3 Q4 to 17 Nov
Commits pushed 3,842 8,538 15,049 21,459
Commits landed 2,634 4,513 8,358 8,787
Pull requests opened 129 2,285 3,935 4,561
Pull requests that merged 94 2,128 3,676 3,425
Lines of code added 514,000 626,000 1,142,000 958,000
Lines of tests added 92,000 278,000 994,000 1,000,000
Test cases at quarter end 4,035 14,073 42,042 76,580
CI runs 13,794 18,675 70,188 87,452
Tickets created 260 1,025 3,568 3,594
People committing in a typical month 8 8 10 12
Commits landed on main branches per month, 2026Bar chart. Jan 958, Feb 715, Mar 961, Apr 1006, May 1799, Jun 1708, Jul 1945, Aug 1414, Sep 4999, Oct 5423, Nov 336402.0k4.0k6.0kJan958FebMarAprMay1.8kJunJulAugSep5.0kOct5.4kNovto 17 Nov3.4k
Commits landed on main branches each month in 2026. November is a part month.

Half a quarter in, the fourth quarter has already landed more commits than the whole of the third. If the rate holds, it will finish at roughly 16,800 landed commits and 6,600 merged pull requests, about twice the third quarter. Landed commits per person committing rose from about 110 a month in the first quarter to about 450 a month now.

The team did not grow fourfold. What changed is in the rest of this post.

Where the effort went

The biggest change this year was not speed. It was focus.

Share of commits Q4 2025 Q1 Q2 Q3 Q4 to date
Event platform 77% 48% 11% 3% 3%
Customer apps (take-back, warehouse, portals) 20% 23% 16% 9% 7%
Physical AI and vision under 1% 14% 1% 4% 5%
Seesa, our on-premises AI product 0% 7% 70% 82% 81%
Site, tooling and internal 2% 7% 2% 1% 4%

A year ago three quarters of our commits went into an event platform. In the first quarter of 2026 the work was spread across four areas. From May it collapsed onto one product, and it has stayed there.

That concentration is a decision, and it has a cost we should name. The customer apps and the vision work did not stop mattering; they got a shrinking share of a much larger total. Commits to the customer apps grew by about half between the first quarter and the third, while the total grew fourfold. We pointed almost all of the new capacity at one thing.

How the work changed shape

Three shifts sit under the totals.

From commits to pull requests. In the first quarter we opened 129 pull requests and landed 2,634 commits. Most work went straight to a main branch. By the third quarter nearly everything went through a pull request with automated checks, and a merge queue landed it. That is what made it safe to let more agents write.

From code to tests. In the first quarter we added one line of test for every five and a half lines of code. In the fourth quarter we are adding slightly more test than code.

Test cases in our main branches at each month endLine chart. 2025-12-31 1678, 2026-01-31 2616, 2026-02-28 3052, 2026-03-31 4035, 2026-04-30 4846, 2026-05-31 9796, 2026-06-30 14073, 2026-07-31 21449, 2026-08-31 27517, 2026-09-30 42042, 2026-10-31 65345, 2026-11-17 76580020k40k60k80kDec1.7kJanFebMar4.0kAprMayJun14kJulAugSep42kOct17 Nov77k
Test cases on our main branches at the end of each month, December 2025 to 17 November 2026.

We had about 1,700 test cases at the end of 2025 and have about 76,600 now. We want to be careful with that number. Agents write tests readily, and a test that an agent wrote to pass its own change proves less than one a person wrote to catch a known failure. The count shows that changes arrive with tests. It does not show that the tests are good, and we have not yet measured that.

From working hours to all hours. In the second quarter 4% of commits were made between ten at night and six in the morning. This quarter it is 14%. Most of that is agents running unattended.

The agents

We write software with coding agents: AI models that are given a ticket, a private copy of the repository and a set of rules, and that open a pull request when they are done. Agents that build our software work on our own code. They do not see customer data, which stays on the customer's hardware.

Through the year, the share of our commits carrying an agent's signature went from 32% in the last quarter of 2025 to 82% in the first quarter of 2026, and has stayed between 77% and 86% since. The real share is higher, because only some of the agents we use sign their commits.

The models changed under us the whole time. The signatures in our history show nine versions of Anthropic's top Claude models in eleven months. We adopted each within weeks of release, and we did not once need to change how we work to do so. The practice (a ticket, a private working copy, tests, a pull request) outlived every model.

Until October we kept no proper record of our agents beyond those signatures. That is the largest gap in this report: for nine months we can say that agents wrote the code but not how many ran, on which model, or how often they failed. We started recording in October, and the detail below comes from one lead engineer's agents on two of our machines. It is a sample, not the whole team.

Agents started per week, by model familyStacked bars. week of 2026-10-05: DeepSeek 117, Claude Sonnet 87, Claude Opus 189, Other 54; week of 2026-10-12: DeepSeek 69, Claude Sonnet 89, Claude Opus 8, Other 273; week of 2026-10-19: DeepSeek 74, Claude Sonnet 27, Claude Opus 7, Other 104; week of 2026-10-26: DeepSeek 925, Claude Sonnet 88, Claude Opus 22, Other 56; week of 2026-11-02: DeepSeek 1088, Claude Sonnet 560, Claude Opus 82, Other 27; week of 2026-11-09: DeepSeek 1205, Claude Sonnet 334, Claude Opus 120, Other 71; week of 2026-11-16: DeepSeek 361, Claude Sonnet 180, Claude Opus 32, Other 49DeepSeekClaude SonnetClaude OpusOther05001.0k1.5k2.0k5 Oct44712 Oct43919 Oct21226 Oct1.1k2 Nov1.8k9 Nov1.7k16 Novpart week622
Agents started each week since 5 October, by model family. "Other" is Codex, GLM, Claude Haiku and open-weight models on our own hardware.

In those six weeks that one seat started 6,298 agents against 1,036 tickets. A third of them were started by another agent, not by a person: a lead agent plans the work, hands each ticket to a worker, and reviews what comes back.

Three patterns stand out.

We were using the most expensive model for everything, by default. In the first week of October the largest Claude model ran 42% of our agents. When we looked, almost none of that was a choice. It was the default setting. We wrote a routing table (which model for which kind of job, with the reason written down) and changed the default. In the most recent full week the largest model ran 7% of agents, a mid-sized Claude model 19%, and DeepSeek, a much cheaper model used through its API, 70%. We landed slightly more per day in November than in October.

Agents read far more than they write. Across the Claude sessions we logged, the models wrote about 69 million tokens and re-read about 19 billion from cache: roughly 280 tokens read for every token written. Seven in ten of their actions were shell commands. Running an agent is mostly paying for it to look at things, so the size of what you make it look at matters more than the length of what it writes.

A growing share of agents do not write code at all. Many of the agents in that count were reading: auditing documentation, sorting through records, checking one system against another. We did not plan for that. It happened because once starting an agent is cheap, reading work that nobody had time for becomes worth doing.

Hardware, and what "local" has meant so far

During the year we brought hardware in-house: pairs of NVIDIA DGX Spark machines for serving models, and a multi-GPU workstation for training, experiments and heavy builds.

Having the machines changed what was cheap. Long experiments that we would have rationed on rented GPUs now queue up and run overnight, and agents start work on the workstation instead of on a laptop. And because our product runs on the same kind of hardware inside our customers' premises, every day we work on these machines is a day spent on the thing we ship.

One result ran against our expectations. We tried open-weight models on our own hardware as coding agents, and they are not yet good enough for anything but small, well-bounded changes: on long tasks they ran out of context and stopped. We ran too few of these to quote a rate. Today our local models serve the product, run evaluations and handle small or sensitive jobs, and the code is still written by larger models in the cloud. Moving more of the writing onto our own machines is a goal for 2027, and this series will track it.

Where it went wrong

We create work faster than we finish it. We raised 3,594 tickets in the first 48 days of this quarter, more than in the whole of the third. About half of them are still open. For tickets raised in the third quarter the figure is 13%. Agents make it cheap to find problems and write them down, and some of our tickets are now raised by agents. Nothing made it equally cheap to decide which ones matter.

Tickets take longer to close, not shorter. The median time from raising a ticket to closing it was a little over two days in the first quarter and about five days this quarter, with a peak near seven in the third. The work inside a ticket got faster. The queue around it got longer.

More work is thrown away. In the second and third quarters about 6% of pull requests were closed without merging. This quarter it is 19%. Part of that is mechanical: when many agents change the same files, pull requests conflict and get replaced. Part of it is real waste, including agents redoing work that had already landed. Pull requests also got bigger, from a median of about 150 changed lines in the second quarter to about 300 now, and bigger changes collide more often.

Most changes are checked by tests, not by a person. Fewer than one in five merged pull requests this year carries a recorded review on GitHub. Changes to our integration branch merge when the automated checks pass; a person reviews before anything is released to a customer. We think that is the right place to spend human attention, but it puts a great deal of weight on the tests, which is exactly the thing we said above we have not yet measured.

The share of fixes has not fallen. About three in ten landed commits this year are fixes, and the figure in the fourth quarter (28%) is no better than in the second (27%). Writing faster has not meant writing more correctly.

We ignored a broken check for nine months. One scheduled sync job in a configuration repository failed on almost every run from January to September, more than 13,000 times, and made up most of our CI failures in the first half of the year. Nobody turned it off. On the main product repository about 6% of CI runs fail this quarter. A check that always fails teaches everyone to stop looking at checks.

We depend heavily on one person. In the second and third quarters one seat accounted for between 43% and 49% of all commits. Agents amplify whoever is directing them, and so far they have amplified our distribution of effort, not flattened it. More of the team is now running agents this way: the number of people committing in a month reached 14 in November, the highest of the year.

What we take from it

  1. The practice matters more than the model. Nine model versions came and went. Tickets, isolated working copies, tests and pull requests carried the gains across all of them.
  2. Look at your defaults. Our most expensive model was doing most of the work because a setting said so. A written routing table cut its share from 42% to 7% while output rose.
  3. Record your agents from the first day. We cannot analyse nine months of our own history because we did not log it. Model, ticket, start, end and outcome would have been enough.
  4. Faster writing moves the bottleneck; it does not remove it. Ours moved to review, to merge conflicts, to CI capacity and to deciding what is worth doing. Plan for those before you add agents.
  5. Count what is thrown away. Commits pushed against commits landed, and pull requests closed without merging, are the cheapest waste measures we have found.
  6. Tests become the reviewer, so test the tests. This is the item we are least far along on.
  7. Bring some hardware in-house if your work allows it. It changed which experiments we were willing to run.

Next month

The next issue will cover the rest of November and December. We intend to add three things we could not report this time: agent records for the whole team, not one seat; a measure of test quality; and the cost per merged change, by model.

Pentatonic Labs · 18 November 2026