Field notes
Specs, agents, and what the measurements actually show
Every claim here is dated and linked to its primary source. When the evidence cuts against the argument, we say so in the piece.
- 03The handoff didn't die. It moved into the repo.Engineers absorbed product decisions; PMs started shipping code. A May 2026 study of 100,000 GitHub developers found coding activity up 180%, releases up 30%.9 min read
- 02A bad plan is worse than no plan at allPlanning is 14% of an agent session and the part you still decide. A 16,991-trajectory study found a subpar plan hurts more than no plan at all.9 min read
- 01Your agent isn't failing. It's guessing.Coding agents don't fail on vague instructions, they guess. A July 2026 benchmark held everything fixed but the wording: wrong-target rate went 9.6% to 75.1%.9 min read