Claude Code, in practice

Nothing is sent without a human go

A hook blocks every command that would send a message to a person (mail, messaging apps) until a human message containing go envoi <name> arms it, for thirty minutes and one message. From July to October, blocked attempts went down each month (59, 55, 42, 33) while sends approved after a go went up (33, 90, 154, 214).

Source: the hook's own log, from 17 July 2026 (first month partial), test sessions excluded. "Lock tampering" = a session trying to create or edit the arming flag itself, always refused.

Fewer blank cheques

In June almost every message ran with permissions bypassed. By October most ran in auto mode, where a classifier reviews each action, or in default mode, where risky actions ask first. Over the same months the volume of work went up.

Source: permission mode recorded on each message in the session logs.

Sessions and sub-agents

Sessions with a human are the ones someone talks to. Scheduled and bot sessions run alone (morning brief, inbox sweep, watchers). Sub-agent sessions are spawned by both to research, review or build in parallel, then report back.

Source: session logs of the main machine, from June 2026 (completed months only). Second chart: sessions open at the same moment, human or scheduled; line = median day of the month, dashes = busiest day.

What the agent actually does

Tool calls, all sessions. MCP groups every connector (mail, calendar, browser, design tools) under one bar.

Models, month by month

Silent failures, written down

The dominant failure in this setup is silence: an automation that did nothing and said nothing. Each one goes into a register with what did not happen, why nobody saw it, and the signal added so it shows next time. The method and anonymised cases are public in proof-gate; the inbox built on the same rule is inboxt.