On September 29, I gave the keynote at AgenticOS 2026, the 2nd Workshop on Operating Systems Design for AI Agents, co-located with SOSP 2026 in Prague. The slides are now online, and I have also archived them on Zenodo so that they can be cited.
Download the slides (PDF) · Zenodo: doi.org/10.5281/zenodo.23187066
The talk opens with three failures. An agent retries an order after a timeout even though the first order already went through, so two orders ship. Another agent loses a standing rule when its context is compacted and then emails every customer without sign-off. A workflow merges a pull request after a rebase by reusing an approval that was given for an earlier version of the code. In each case the model made a reasonable decision and the outcome was still wrong, because nobody had written the rule down in a form the runtime could check.
This is what I mean by the harness wars: we are building operating systems for agents without having agreed on what they guarantee. The talk proposes four contracts that an agent runtime should state and test, namely effect (when an action becomes real), state (what must survive compaction), composition (what survives when steps and agents are chained), and intent (who checks the code a model leaves behind). For each one, I go through what we can already state today, what existing systems provide, and which parts remain open.
Thanks to the AgenticOS organizers for the invitation.
Cite
Laurent Bindschaedler. The Harness Wars: Why Agents Need a POSIX Before a Kernel. Keynote, AgenticOS @ SOSP 2026, Prague, September 2026. https://doi.org/10.5281/zenodo.23187066
@misc{bindschaedler2026harnesswars,
author = {Bindschaedler, Laurent},
title = {The Harness Wars: Why Agents Need a {POSIX} Before a Kernel},
howpublished = {Keynote, 2nd Workshop on Operating Systems Design for AI Agents (AgenticOS @ SOSP 2026), Prague},
month = sep,
year = {2026},
doi = {10.5281/zenodo.23187066}
}