MLOps.community The Hidden Bottlenecks Slowing Down AI Agents
93 snips
Aug 5, 2025 
Paul van der Boor, VP of AI at Prosus Group, and Bruce Martens, an AI Engineer at the same company, dive into the real bottlenecks in AI agent development. They reveal that it's not just about the tools but also the importance of evaluation and feedback. The duo discusses innovative strategies for creating effective evaluation sets and the impact of existing AI tools on productivity. They also examine the complexities of building custom AI solutions and the challenges of navigating cognitive load when reviewing code, highlighting the balance between AI assistance and developer mental strain.
AI Snips
Chapters
Transcript
Episode notes
Buy Tools When Possible
- Encourage teams to buy existing tools to avoid wasting time building what already exists.
- Focus internal resources on core product areas where off-the-shelf solutions don’t fit or lack maturity.
User Experience Drives Tool Choice
- Teams with deep AI expertise need more low-level, headless tools rather than drag-and-drop solutions.
- Different users require vastly different tooling experiences, from engineers needing SDKs to non-technical users wanting simple interfaces.
Coding Agents Aid Incremental Dev
- Paul’s team uses coding agents like Devin and Cursor for incremental changes on well-documented codebases.
- They track usage with KPIs such as percent of pull requests done by agents to measure adoption and effectiveness.