Tags

The engineering substrate of production agents: context handling at scale, skill reuse, self-modifying code, multi-agent coordination, and the interference and failure modes that surface when components are stacked. A mix of research summaries and release coverage.

Tracing, metrics, and cost attribution for agent systems in production: wiring OpenTelemetry through LangGraph, vLLM, CrewAI, and LightRAG pipelines, alongside coverage of Grafana's AI-observability tooling and the o11y-bench benchmark. The throughline is making an agent's tool calls, token spend, and failure modes legible after the fact.

Coverage of LangGraph, the graph-structured agent runtime from LangChain: release notes (node error handlers, streaming, the TypeScript port) and end-to-end tutorials that build retrieval, PR-review, and traced agents on top of it.

Post-training and reinforcement-learning methods aimed at agents: RLVR (reinforcement learning from verifiable rewards), skill generation and curation, task routing across frozen expert models, and closed-loop training pipelines. Mostly research summaries drawn from the arxiv wave.