Agent observability with OpenTelemetry
A reference harness (reasoning loop, memory, evals, guardrails, human in the loop) with run-level tracing across models and tools. The metric that matters: cost per successful task, because a cheap model that keeps retrying can cost more.
Least-privilege MCP for production data
The real security boundary is the database account the MCP server logs in with. Give each agent its own identity, cap results (25 rows default, 500 max), and move recurring questions to reviewed, parameterised queries.
Agent costs are collapsing
Hex's cost per agent conversation fell from $6.44 to $1.64 this year. On Vercel's AI Gateway, small "decision models" went from 0 to 26% of requests between August and October.
Agents in production, owners unclear
Retool finds 89% of IT leaders have at least one agent in production, with no clear standard for who owns them.
Products go headless via MCP
Anthropic's Interactive connectors let Adobe, Atlassian, Notion and Box ship their full UI inside Claude; Granola's MCP 2.0 search uses 85% fewer tokens.
A RAG onboarding agent for marketplace partners
Booking.com's ML team on replacing static help content with a retrieval-based onboarding agent for partners.
Deep dives
RoPE and positional embeddings in 2026 open-weight LLMs, and why agent-driven rewrites still need humans: code is cheap, verification isn't.