API Platforms For Scale

The episode covers substantive API governance topics including the three-pillar approach (people, processes, tools), contract testing, Arazzo for multi-turn workflows, and AI-readiness of API specifications. However, significant portions involve biographical background, small talk, and general encouragements to 'listen to people' that dilute insight density. The specific technical insights (e.g., accuracy improvements from 70-80% to 99.6% with proper documentation, or Arazzo boosting multi-turn accuracy from 60% to 90%) are valuable but scattered amid conversational padding.

“putting descriptions, using standards, examples, that bumps the accuracy to almost 100%. That's like 99.6 or something like that, from our test at least.”

“if you describe this workflow a deterministic way using something like Aradzo, the accuracy goes up to over 90%.”

While the core frameworks (API governance as people-processes-tools, spectral linting, contract testing) are standard industry practice, the guest offers some original angles: connecting AI agent accuracy improvements to API specification rigor with concrete benchmarks, using Arazzo for deterministic workflows rather than hypermedia, and the pragmatic 'minimal viable standards' approach. However, these insights are not deeply contrarian - they represent thoughtful execution of known patterns rather than fresh paradigms.

“if you follow these standards, AI agent figure out that it's a standard. If you have an open API file and give descriptions of what API is for... that will give probably enough context to the LLM it understand what has to do.”

“Hypermedia was not adopted because it actually makes design of much more complex... And agents actually work for understanding hypermedia because they work very well with dynamic text right? And Aradzo is solution similar... it's simpler, I think.”

Andrzej Jarzyna is a highly credible practitioner with 15+ years in API governance across multiple tier-1 companies: 3scale (API management pioneer), Adidas (led enterprise API governance program, published early public API standards), ING Bank, and now PZU as Chief API Architect. He has authored a published book and speaks at major conferences. He brings genuine operational experience at scale (70 teams, 1,300 IT staff at PZU), not theoretical positioning. This is exactly the caliber of guest who has actually executed governance at enterprise scope.

“I started in 2011 at of three first API companies, Threescale... I later started a true API governance journey at Adidas where we had already an API governance team and we're leading the program for the whole company... I worked at ING bank, helping them with their API governance program”

“So we built a spectral rule set... and put it into CI-CD pipelines... we also built a plugin for Visual Studio Code and IntelliJ.”

The episode includes some concrete specifics: 70 product teams and 1,300 IT staff at PZU; 99.6% vs 70-80% accuracy from tests; Arazzo boosting multi-turn accuracy from 60% to 90%; requirement for OpenAPI v3, descriptions, and naming conventions only (not over-restrictive). However, large sections remain vague: no specific metrics on adoption rates, timeline specifics are fuzzy ('a bit more than a year and a half ago'), and many claims lack supporting numbers. The 'breaking changes' process and contract repository are mentioned but not quantified.

“70 teams in the company, 70 product teams and around 1300 people in IT”

“when you do a pull request to your repository, there is a process which verifies or validates your open API file against our spectral rule set. If it fails, well, it fails merge or pull request.”

The host asks reasonable opening questions and follows up on key topics (AI readiness, tooling, standards evolution), but rarely pushes back or probes deeper. When the guest offers vague claims ('it went slower than I anticipated'), the host moves on rather than asking for specifics. The conversation is collegial but lacks the sharp questioning that would stress-test claims. Few genuine disagreements or productive tensions are explored. The host is competent but safe, functioning more as a guided tour than investigative dialogue.

“Tell us the story. So how did it go?”

“So when all the craze about AI started, I was kind of... I wouldn't say afraid, but worried a little bit”