The period treats large language model (LLM) agents as trainable and inspectable software systems. EvoTrainer, FLARE, and SPOQ carry the strongest evidence: better agents come from diagnostics, gated execution, and task…
Agentic coding work is becoming easier to improve when teams preserve intermediate evidence: suspicious lines in repair loops, explicit dependency gates in multi-agent execution, and trace quality signals for…
Current emphasis: coding agents are judged by how they run, repair, and stay inside bounds. A-ProS shows gains from stateful judge feedback. ProcBench scores process defects inside traces.
Coding-agent teams can test runtime scope, trace quality, and file selection as separate parts of the execution path.
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…