Autonomous agents

Bots that work on their own

How xAI, OpenAI, Anthropic, Microsoft and Google govern their autonomous agents today, and who answers for what they do.

Ricardo Martínez · Oct 26, 2026 · Editorial summary v1

Read PDF · EN ↗Leer PDF · ES ↗

Bot builders already know how to trust what bots deliver. Companies that use them still do not know how to trust what bots do.

The two identity models, the gaps the vendors themselves admit and what the 2026 data says. A 75-second video, Spanish audio with English subtitles.

Key data

What the documentation and the data say

What it adds

It reviews the official documentation of five vendors, current as of October 26, 2026, with five questions: what identity the bot acts with, what separates it from other bots, what gets recorded, what is on by default, and who is accountable. It finds two models: in the most common one the bot acts as the user and has no identity of its own; only Microsoft and Google give it one, with a human sponsor and permissions that expire. It starts from Lauren Tan's method for trusting what a bot delivers and carries it into governance, with seven steps for a company that already runs bots.

Method and limits

It is a documentation review, not a product test: nothing was tested live. Every source is labeled primary, secondary, interested party or own observation. The surveys cited are vendor-funded and publish no margin of error, and widely repeated figures without a primary source were left out. Products change weekly; the cutoff is October 26, 2026.