Hermes Wiki
opportunities/2026-08-15-incremental-catalyst2026

Opportunities: Wiki Knowledge Applicable to the Catalyst 2026 Talk — Incremental Pass, 2026-08-15

Scope: this pass checks Claude Code's Auto Mode default (this window) against the talk's central "propose-only, human-gated approval" design decision (§3–4.2 of Catalyst2026-Toronto-Network-That-Answers-Back.md).

The talk's approval-gate design assumes human review is a reliable safety check — this window's Auto Mode data says that assumption needs qualifying, not abandoning

The talk's architecture puts a human approval click between every agent-drafted change and the digital twin (§3, §4.2: "a human approves the intent of the draft before any compute is spent validating it"). Anthropic's own study behind Claude Code's new Auto Mode default is a direct, dated data point that complicates the unqualified version of that design: across 1,053 paid testers, humans manually approving every action caught only 13.6% of genuinely harmful actions, because in practice people approve 97% of prompts regardless of content — a classifier trained specifically on irreversible/destructive/external-facing actions caught 89% instead. The mechanism named for the gap is prompt fatigue: a human whose attention has already been worn down by the two-hundredth approval click of the day isn't meaningfully reviewing the two-hundred-and-first.

This doesn't invalidate the talk's design — it sharpens the argument the talk should make on stage. The talk's approval gate isn't a raw "click yes/no" prompt; it's gating proposed network changes specifically, which is exactly the class of action Anthropic's own classifier is trained to flag as high-stakes (irreversible, externally visible) rather than the low-stakes majority of actions a general coding agent takes. But the 97%-rubber-stamp finding is still a real risk to name explicitly for a P1/P2-frequency approval flow: if the network agent proposes changes often enough that approval becomes routine, the same fatigue effect Anthropic measured will degrade the human review step's real safety value over time, not just in theory.

Concrete next step: add this as a Q&A prep line (§9-adjacent) for the "isn't a human approval click just theater" question, which a 301 technical audience is likely to ask given this exact debate is live in the industry right now: cite the 97%-rubber-stamp/13.6%-catch-rate finding directly, then state the mitigation the talk's own design already has that a generic approval prompt doesn't — the digital-twin validation step after approval catches exactly the class of error a fatigued human reviewer would miss (drift between the approved intent and current live state), which is a second, mechanical check that doesn't depend on human attention the way a bare approval click does. That's a stronger answer than either dismissing the concern or claiming the approval click alone is sufficient.

Hermes Wiki