How Poor API Architecture Makes AI Agents Expensive - Sebastian Małyska

📘 Free e-book: The 7 success factors of software testing. 25 years of project experience in one 33-page workbook, now also in English 👉 Get it for free

"Improving is not like giving you know golden card to AWS, give me more resources. But this is the hard way, this is engineering way." - Sebastian Małyska

With Sebastian Malyska I talk about what sits behind the APIs testers often take for granted. We get into why REST is still on top but not the whole story, how GraphQL and webhooks fit into the picture, and what actually happens when an AI agent talks to a real system through MCP. There is a practical argument running through all of it: if you understand the protocol layer, you can make smarter decisions about where to test, how to structure automation, and how to avoid burning unnecessary tokens and budget. Sebastian also makes the case that front-end testers are already closer to this world than they think, and that the next step in the career is simply to look one layer down.

Sebastian Małyska is a software quality enthusiast with over 20 years of experience proving that “it works on my machine” is not a testing strategy. Has worked as a manual tester, automation engineer, QA Lead, and QA Manager, basically wherever quality (or patience) was missing. Speaker at national and international conferences and a program committee member who enjoys talking about QA almost as much as reporting other people’s bugs. Codes mainly in Python, because life is too short for bad code and even worse languages. President of the Polish Quality Board, Secretary of the ISTQB® Board, member of the IREB® community, and co-founder of ŁuczniczQA in Bydgoszcz. On a daily basis, fights for better quality, fewer critical defects, and more common sense in IT projects.

Highlights:

• APIs optimized for agent use require fewer requests to retrieve the same data, which directly reduces token consumption and lowers operational costs.
• REST is a set of recommendations, not a strict protocol, so teams can and do deviate from its conventions without violating a formal standard.
• MCP works by having the MCP server send a description of its available tools and resources to the LLM at the start of a session, which is how the model learns what actions it can call.
• API-level tests offer greater stability and faster execution than end-to-end browser tests, and the right testing strategy names the layer where each scenario is best covered rather than defaulting to front-end automation.
• Starting API test design from documentation such as a Swagger spec, before the front end exists, removes the wait time at the end of the development cycle and eliminates idle waste.