The strongest signal is that agents need inspectable execution and stricter task evidence. DuST uses execution-labeled candidate code as training data, Shepherd records live agent state for branching, and ComplexMCP…
Agent teams can move faster by adding trace, state, and provenance checks around the places where agents already fail: coding runs, MCP tool workflows, and automation workflows that mix untrusted text with secrets or…
The day’s strongest theme is control around LLM coding systems. New work tries to cut prompts to the code that matters, trace what crosses the natural-language/program boundary, and fence agent actions with tighter…
Concrete work is forming around three control points in LLM coding systems: shrinking repository context before repair, restoring taint and slicing across LLM call boundaries, and limiting what coding agents can touch…