Patange takes a bare loop (call a model, use tools, repeat until done) and grows it into a harness, introducing context management, a filesystem for intermediate work, tool boundaries, persistent memory, and subagent delegation one problem at a time. The material comes from Deep Agents, an open-source model-agnostic harness on LangGraph.

This is the closest thing on the schedule to a second opinion on the harness question. The interesting part is the order: which pressure forces which component, and what each one costs. Worth comparing against what the browser-harness work here found, namely that the harness itself is the thing most likely to fail silently.

Herrmann's claim is that technical literacy is a governance control, not a prerequisite to one, and that "trust the vendor" and "trust the eval score" fail for the same reason: nobody in the room can say what the system does when it is wrong. She promises examples of risks caught or missed on exactly that basis, and an argument that deterministic and neurosymbolic techniques give more explainable failure points than an end-to-end LLM pipeline.

"Trust the eval score" failing without someone who can explain the mechanics is the same finding as an uncalibrated gate, arrived at from the governance side rather than the testing side. See trust boundaries as a typed category.

Taking a prompt-generated prototype and making it secure, scalable and affordable: exposed API keys, brittleness on a malformed model response, unbounded spend. Adjacent to elenctic vibe code review, which asks what review even means when nobody wrote the code.

The keynote is a career-framing argument (the "can you build it?" question is answered by machines; "should this exist, and how would we know if it were good?" replaces it). Isler covers product management under AI. Voulet demos Antigravity 2.0 against Google Cloud and Workspace MCPs, which is worth seeing purely for how the MCP auth story is presented. Mukherjee drives Blender with Gemma 4 agents on Hugging Face Spaces.