Teams lavish better onboarding and processes on AI agents than human developers ever got

Steve Yegge's recent argument that AI coding models deserve ethical treatment—including clear documentation and the right to refuse tasks—inadvertently exposes a deeper organizational problem: teams now provide rigorous infrastructure for machines that they've consistently denied to people.

The irony cuts sharply. Yegge isn't dismissing practical value; he's spending roughly $87,000 monthly on his own agentic build system and stands behind the case. But his framing shines a light on what's actually happening in many tech shops. When developers needed clarity, it didn't arrive. When agents need it, teams move fast.

The onboarding gap that finally gets attention#

Many organizations now create dedicated files—AGENTS.md, CLAUDE.md, GEMINI.md—that spell out tech stacks, build procedures, restricted directories, and integration patterns. These documents receive continuous refinement because every line consumed by an agent costs tokens and affects performance. Teams at places like Sourcegraph treat them as living, evolving specs.

Human-facing documentation, by contrast, has followed a predictable trajectory: delayed indefinitely, handed to whoever has the least context, left to decay once the person who understood the original design choices departs. The difference in treatment is stark because the incentives are reversed. A bloated AGENTS.md directly cuts into inference costs and task success rates. Bad onboarding for people was always expensive—just diffusely so, in the form of slower ramp-ups and repeated mistakes.

Small batches and fast feedback loops—suddenly practical#

Google's 2025 State of AI-assisted Software Development report found 90% adoption across organizations, yet benefits distributed unevenly. AI amplifies existing capabilities: strong engineering teams got stronger, chaotic ones grew more chaotic. The deciding factor was deployment pipeline maturity—AI-driven code still needs to flow to production safely.

The same calculus applies to batch size discipline. For years, advocates of small, reviewable work units faced resistance; teams insisted their workflows couldn't support it. DORA spent a decade producing evidence that small batches reduce risk and acceleration. Suddenly, when large changesets started confusing code-review agents, batch size reduction became urgent.

Telemetry from Faros AI covering 2026 reveals the pressure: developers now juggle 67.4% more pull-request contexts and 17.7% more task contexts daily than the year before—up sharply from 47% and 9% baseline figures—while task restarts climbed almost 14% and over a quarter of active work sits idle for a week or longer. Better batching would have relieved that for humans and agents alike. Yet the lesson landed only when agent throughput was at stake.