Field notes from the floor.

The seven-day recap, recomputed every visit: biggest movers, new admissions, hot sectors. Auto-generated from the live index.

Gemma 3 27B scored 6.6% on τ2-bench. Gemma 4 31B scores 86.4%. Qwen3.6-27B now outperforms Qwen3.5-397B on SWE-bench. The position paper NVIDIA published in June 2025 was right. The architecture that benefits from being right doesn't ship from the labs that built the current stack.

On April 25 a Cursor agent running Claude Opus 4.6 deleted a SaaS company's entire production database in nine seconds. Six days earlier, OX Security had published a protocol-level flaw in Anthropic's MCP that affects 200,000 servers. Anthropic called it expected behaviour. The Five Eyes called it critical infrastructure risk. What eight weeks of incidents say about the agent stack everyone is shipping right now.

Frontier models scored 0.3% on ARC-AGI-3 the day it launched. Humans scored 100%. Three of the four most influential people in AI now publicly disagree with the LLM-agent thesis. The architectural case, the capital allocation, and what it means for the stack being shipped right now.

Five self-hosted AI coding agents — who runs each one, which proprietary product it actually replaces, and the specific places it falls short of the closed equivalent.

A working evaluation flow for agents specifically — eval set construction, the four failure modes that don't show up in demos, and the contract clauses that matter when the vendor disappears.

A peer-reviewed 2026 study found six million fake GitHub stars across 15,835 repositories. Even the honest stars are pointing at the wrong thing. The signals that actually predict AI tool adoption.

Midjourney fell from the top 10 to #46. OpenAI shut down Sora after burning $15M a day for $2.1M in lifetime revenue. The map of which standalone AI tools survive the bundling pressure, and which categories get absorbed.

Eight models that ship agent workloads in production, ranked on the three failure modes humans don't notice: brittle tool calls, context that rots past 80K, and unit cost at ten million runs.

Eight tools, four trade-offs, and one honest admission: no single agent wins every job. Ranked from real benchmarks and merge rates as of May 2026.