smallcode's 87% benchmark with 4B models shows efficiency gains that could reshape our infrastructure costs. When small models deliver frontier-level coding performance, the economics change dramatically.

The 47-point swing in benchmarks depending on judge choice exposes a fundamental flaw in how we measure AI progress. Single-judge evals are measuring preference, not capability.

Changes how we think about development context. We can now review and adjust Copilot suggestions from anywhere, which could be useful for code reviews or quick fixes whilst away from the desk. Depends on workflow habits.

Essential reading for agent evaluation strategy. Covers the difference between testing model capabilities versus agent performance, with practical frameworks we can implement. Particularly useful for production deployments.

Between 12-factor-agents principles and Karpathy's Claude optimisations, we're seeing battle-tested patterns codified. The experimental phase is ending; the engineering phase begins.

Practical pattern for adding persistent memory to CLI agents. Shows exactly how to wire up MCP servers with cloud memory services. Useful if we're building command-line tools that need to remember context across sessions.

Framework milestone from a major AI vendor. Read the release notes and decide if it changes how we build agents or wire tools into our stack.

Crucial data for infrastructure decisions. If we're scaling coding agents, these benchmarks show where to optimise for cost versus speed. The TTFT numbers are particularly relevant for interactive use cases.

Excellent primer if we need to catch stakeholders up on LLM progress. Simon's consistently good at cutting through hype to focus on what actually matters for practitioners. Great for team education.

Best practical introduction to agent architecture we've seen. Uses familiar Logo programming to show how voice input, model reasoning, and tool execution connect. Perfect for explaining agents to stakeholders or new team members.

Important perspective for anyone managing dev teams. Individual juniors who expand their scope will thrive, but we need to preserve mentorship structures. Don't just pair juniors with AI; pair them with experienced engineers too.

Critical if we're using LLM judges for evals. Always run multiple judges and average the scores, or use the same judge as the model we'll deploy. Single-judge evals are measuring preference as much as capability.