Until June 2025, one engineer maintained one SDK by hand. Then we went agentic. That engineer scaled up, and then the approach scaled out to the people around him. Here is what happened, measured from the git history of the 22 repositories it grew into, comparing the 12 months before with the last 12.
From late 2023 to May 2025, the Camunda 8 JavaScript SDK (camunda-8-js-sdk) was maintained manually. It was a healthy, well-used project, and it moved at the speed of one person with a day job.
Every feature, test and release went through the same pair of hands.
Monthly activity across every repository in this story, coloured by what it was for. The dashed line is June 2025, when the work went agentic.
Agents generate from what you give them. So before generating clients, we made the OpenAPI specification say what the API means: domain types for keys, eventual consistency marked on every endpoint that has it, and lint rules to keep it that way.
That is derivation over duplication applied to a whole product line: one source of truth, everything else generated from it.
camunda/camunda.The Orchestration Cluster API SDKs started with JavaScript. Once the spec, the generator pipeline and the test harness were proven there, each new language reused them. The gaps between languages shrank from months to weeks.
Month of first commit in each orchestration-cluster-api-* repository. The three oldest SDKs have shipped more than 100 releases each.
Seven SDKs at agentic speed only work if “green” means something. So the tests were derived from the spec too: response bodies asserted against the schema, then whole test suites generated from a semantic graph of the API.
The test generator became the most collaborative repository in the story: ten people have committed to it.
Then the SDKs were tested against each other. A conformance harness wraps all seven SDKs in thin adapters, sends the same operation to each, and compares the HTTP requests they would send. Seven SDKs must agree on every operation in the spec.
Customers ask how to deploy their workers: how many processes, how many workers in each, REST or gRPC streaming, which SDK. Before, the honest answer was “profile your workload”. Nobody had the time to measure every combination for even one SDK.
At agentic speed, the measurement is the cheap part. A performance matrix runs every valid combination of worker topology, SDK language, SDK mode, handler type and cluster size, locally or on Google Cloud, and publishes the results to a dashboard.
The answers are now specific. Prefer streaming when you run many workers: its tail latency stays flat while REST long-polling climbs. Spread workers across processes, about one event loop per core, instead of adding workers to one process. Set the C# poll interval to 25 ms: in the Google Cloud runs, that alone took throughput from 117 to 146 jobs a second.
The matrix also feeds the SDKs: it produced the proposal for threaded workers in the JavaScript SDK. For the method behind it, and how the conformance harness works as an oracle, see Point them at a number.
An approach that only works for the person who invented it is a trick. This one transferred: to a student who learned it from scratch, and to a team that took a whole repository over.
She joined as a student intern on work experience and learned agentic engineering from scratch. She is now an engineer on staff.
870 commits and 167k authored lines across 9 repositories in eight months. She is now a co-owner of the SDKs and the principal driver of several of them.
api-test-generator was built in April and May. In June it changed hands. Since July, every commit has come from the QA team and its contributors.
Before going agentic, one engineer owned the SDK. Now the SDKs have an SDK pod of three people. The former intern co-owns them and drives several. A third engineer brings the human bus factor down, and supports the SDKs in production with an agent that picks up production issues, so a report from the field becomes a fix without pulling the others off their work.
The nanobpm organisation started as research spikes. It became a Rust BPMN engine compatible with Camunda 8, compiled to WebAssembly to drive interactive examples in the docs and on the website, plus the orchestration we now use internally for the agentic SDLC.
The biggest spike on the chart, August 2026, is mostly this garden.
Source: the full git history of the 22 public repositories named in this chapter, merges excluded. Bot authors (Renovate, release automation) are excluded from commit and line counts. “Authored” means written by a person or by an agent working for them. Snapshot taken October 2026. The conformance harness and the performance matrix are internal repositories and are not counted in these numbers.
Engineer A hardened the spec first, derived the SDKs and the tests from it, used the new capacity to measure what nobody could measure before, then taught the approach to an intern and a QA team. One owner became a pod of three.
The speed came from the agents. Where it went came from the structure. Two more engineers got there on their own, in different settings: see Three shapes. And in c8ctl, the method left the SDKs.