From an architect standpoint, gen AI is terrible at answers but great at questions.
1. The Thoughtworks Technology Podcast previews Technology Radar Vol. 33, recorded one week before release and focused on its four themes, all about AI.
2. Host Neil Ford, who writes the first draft of each theme, joins Brian Oliver, a principal engineer at Thoughtworks writing a book with O'Reilly on delivery algorithms and owner of the first theme.
3. The radar team narrowed about 300 blips to roughly 100 at a meeting in Bucharest, and the four themes are infrastructure orchestration, agents elevated by MCP, AI coding workflows, and emerging AI anti-patterns.
4. Oliver says Slurm, dating to 2004, is the de facto HPC scheduler, while Kubernetes projects such as Kueue, Kubeflow and Volcano add topology awareness, using rack and network-speed labels so training jobs land on adjacent nodes, with one GB200 rack holding 14 terabytes of unified memory.
5. Ford says the radar predicted a hotspot of AI innovation each cycle, that RAG dominated the last radar and agents and MCP dominated this one, and that context engineering patterns such as agents.md describe each agent's role and responsibilities.
6. The first malicious MCP server appeared in the wild on the Thursday of the meeting, and Lily Ryan of Thoughtworks infosec has seen authentication and authorization implemented at the wrong level in the middle layer.
7. On AI coding workflows, Oliver cites AI plugged into pipelines and Kubernetes clusters, Ford cites results with AI for understanding legacy code bases and tools like v0 and Bolt for UI prototyping, and the hosts advise dropping week-old LLM threads in favor of newer models.
8. A radar attendee asked why the context window wasn't just called "attention span," and Ford describes generative AI as a tool for automating "yak shaving" for experienced people, including the radar designer's added translucency.
9. In the anti-patterns theme, Ford says text-to-SQL looked promising but proved "more trouble than it's worth," and Oliver says spec-driven development pushes teams toward waterfall-like upfront specification and too many files to maintain.
10. This episode suits engineers, architects and technology leaders tracking AI tooling trends.
