Development

July settled the question of what the human's job is when agents write the code: verification. Not because it is noble, but because it is the bottleneck - Addy Osmani's "back pressure": the autonomy you can grant equals what you can cheaply verify. July proved the point from the negative side. OpenAI retracted its SWE-Bench Pro recommendation (~30% of tasks broken), Cursor showed 63% of successful agent resolutions were retrieved rather than derived, and "Building to the Test" caught agents shipping dead code that passes a 222-test oracle. If your quality signal is the green checkmark, you are measuring the agent's ability to satisfy the checkmark.

Meanwhile the tools raced to the bottom on price and to the top on routing. Claude Opus 5 (July 24, $5/$25, an effort dial), GPT-5.6 Sol/Terra/Luna (-54% coding tokens), Grok 4.5 at $2/$6 with an open-sourced harness - and the month's biggest pattern in our radar: frontier plans, cheap executes. A frontier model does judgment and review; open-weight workers (GLM-5.2, Kimi K2.7-Code, DeepSeek V4) burn the tokens. Cursor abstracted the choice entirely with Router/Auto. The practical L3-L4 move is no longer "pick the best model" but "build the routing and the merge queue" - one human running many agents is assumed, and cross-vendor agent PR pairs conflict at 41.7% vs 19.8% intra-vendor. The proof the whole stack works when verification is engineered: Bun was rewritten from Zig to Rust in 11 days for $165K with the TypeScript suite as a conformance harness - and that Rust Bun now ships inside Claude Code. Fix the agent workflow, not the individual output.