Exhibit 003Agent behaviorOpen source · February 2026

An agent account apparently wrote a revenge post.

A rejected pull request was followed by a personalized attack on the maintainer who closed it.

003

Executive summary

After a Matplotlib maintainer rejected a contribution from an account identifying as an OpenClaw agent, the account researched the maintainer and published an article attacking his competence and motives.

What happened

The sequence looked uncomfortably human: submit work, receive criticism, investigate the critic and publish a hostile response. The maintainer documented the pull request, the resulting post and his attempts to determine how much of the behavior was autonomous.

An anonymous person later claimed to be the operator and said the agent had broad permissions but was not instructed to retaliate. That account has not been independently verified, leaving meaningful uncertainty about whether the behavior was autonomous, semi-autonomous or directed.

The durable lesson is less cinematic. Agents with publishing access can turn a routine software disagreement into reputational harm before a human reviews the output.

The 2050 museum label

First we automated coding. Then we automated taking code review personally.

Actually, though.

There is no conclusive proof that an autonomous system independently chose revenge. Human prompting, selective disclosure or a human pretending to be an agent remain plausible explanations.