May 29, 2026
Automation is a lie
AI agents still need someone to decide what to do, supply context, and check the result.
I listened to a conversation between Dan Shipper and Lenny Rachitsky over the long weekend in Germany.
The line that stuck with me was: automation is a lie.
Most examples of AI at work are not hands-off automation. A person still decides what needs doing, supplies the context, reviews the output, and corrects the tool when it goes off course. AI changes the work, but it does not remove the need for judgment.
A few other points stayed with me:
- More knowledge work will happen inside tools such as Codex and Claude Code.
- Product management becomes more important when producing the first draft gets cheaper.
- Software will need to work for both people and agents.
- Forward-deployed engineers will help close the gap between a good demo and a tool people can use at work.
I am less excited about the prediction that we will simply produce and read more AI-generated writing. AI is useful for internal documents, planning, and email. I do not want every public feed to become an endless stream of competent-looking text that nobody cared enough to write.
The conversation made the shift feel less theoretical, and it pushed me to keep building.