Engineering notes from the field.
Essays and field notes on AI, software engineering, design, and the craft of building product teams that ship. Written by the engineers doing the work.
Meta Closes Muse Spark, Freezes Llama's Frontier Track
Meta shipped closed Muse Spark on April 8 and moved frontier AI to Superintelligence Labs — the FY27 open-weights procurement anchor breaks.
Enterprise AI Agents: 80% Embed, Only 31% Reach Production
80% of Q1 2026 enterprise apps ship an AI agent, but only 31% run one in production. 88% of pilots die on eval and success criteria, not on model quality.
Claude Sonnet 5 Ships Cheaper Agents: $2/M, 1M Context, 63% Coding
Anthropic shipped Sonnet 5 on June 30: $2/$10 per M through Aug 31, 63.2% agentic coding, native 1M context — the mid-tier is the new default-route substrate.
MiniMax M3 Lands Open-Weight Frontier Coding at 1M Context
On June 1, 2026 MiniMax shipped M3 — frontier-tier open-weight coding, 1M-token context, 59% on SWE-Bench Pro, native multimodality.
xAI Ships Grok Build /goal Autonomous Mode and Agent Dashboard
On June 24, 2026 xAI added /goal to Grok Build — a long-running autonomous mode plus a single-screen Agent Dashboard for managing parallel coding sessions.
Cognition Ships FrontierCode: Mergeability, Not Correctness
On June 8, 2026 Cognition launched FrontierCode — Claude Opus 4.8 leads at 13.4% on the Diamond split that grades 'would a maintainer merge this?'