Simon Willison reports that an OpenAI desktop app associated with ChatGPT/Codex includes a roughly 1.7GB LibreOffice runtime, apparently to manipulate office documents locally. The supplied snippets establish that OpenAI’s desktop tooling can access local files and desktop applications with permission, but they do not independently confirm the LibreOffice bundle, its exact contents, or whether OpenAI intends it as a general reusable execution layer for document-operating agents.
OpenAI’s reported bundling of LibreOffice is a consequential implementation of Scott’s “give the agent a workshop” and Code-First Architecture positions: system capability comes from packaging a capable execution substrate around the model, potentially letting document work run through a deterministic, machine-operable runtime rather than GUI imitation alone. It offers a dated-receipts opportunity, but the evidence does not yet establish how Codex invokes LibreOffice, whether execution is contained, or whether the runtime is exposed as a reusable general layer.
ip:framework.code-first-architectureip:source.give-the-agent-a-workshop-ebookip:concept.application-ai-fitnessradar:concept.agent-runtimeradar:concept.document-parsingradar:concept.computer-use-agents
queries asked of Scott's wikis
- local execution layers for AI agents
- document-operating agents and office workflows
- bundled runtimes versus native application control
- agent tool packaging and reusable skills
- local file access security and permission boundaries
- headless document conversion in agent harnesses
2026-09-07T14:34:45Z
The credible reconstructed Willison report remains a useful example of packaging document skills with a capable local runtime, but the episode has produced no further technical evidence or expected near-term confirmation. Expire active monitoring without treating the bundle finding as disproved or its broader reusability as established.
2026-09-05T14:23:35Z
Willison’s reconstructed firsthand report describes a multi-tool runtime plus document-plugin skills, supporting a concrete agent-tool packaging finding rather than merely a large dependency. No new evidence arrived at this review; general reusability, invocation mechanics and permission boundaries remain unestablished.
2026-09-03T13:25:42Z
The refreshed discussion is further amplification of known concerns about dependency weight, fidelity and licensing, not a new technical line of evidence. Simon Willison’s firsthand package inspection remains credible, but the broader reusable document-agent runtime interpretation is still uncorroborated.
2026-09-02T12:41:25Z
The refreshed discussion again adds no independent inspection, invocation details, or first-party response; it is repetitive amplification of known concerns. The bundle finding remains credible, but its interpretation as a reusable document-agent execution layer remains uncorroborated.
2026-09-02T11:37:29Z
The refreshed comments remain reactions to dependency size, document fidelity and licensing, without an independent inspection, invocation details or an OpenAI response. Simon Willison’s firsthand package finding remains credible, but the interpretation of the bundle as a reusable document-agent execution layer is still uncorroborated.
2026-09-02T10:30:26Z
The latest discussion remains repetitive reaction to the known bundle, without independent inspection, invocation details, or an OpenAI response. The firsthand package finding is credible, but the broader claim that it establishes a reusable document-agent execution layer remains uncorroborated.
2026-09-02T09:37:22Z
The refreshed discussion remains repetitive commentary on bundle weight, fidelity and licensing, without independent inspection or new details about how Codex invokes the packaged tools. The firsthand bundle finding remains credible, but its interpretation as a reusable document-agent runtime is still uncorroborated.
2026-09-02T08:33:04Z
The refreshed comments again add only reactions about quality, dependency weight, licensing and competitive implications, without independent inspection or invocation details. The credible firsthand bundle finding stands, but there is still no support for promoting the broader reusable-runtime interpretation.
2026-09-02T07:36:55Z
The refreshed discussion adds no independent inspection, invocation details, or first-party response; it remains repetitive commentary about dependency weight, quality, and licensing. The firsthand bundle finding is credible, but the broader reusable agent-runtime interpretation remains uncorroborated.
2026-09-02T06:25:18Z
The refreshed comments remain speculative reactions about dependency weight, quality, licensing, and competitive implications; they add no independent inspection or details of how the document skills invoke LibreOffice. The broader reusable agent-runtime interpretation therefore remains plausible but uncorroborated.
2026-09-02T05:32:29Z
The refreshed comments add no independent inspection, invocation details, or first-party response; they continue to amplify known concerns about size, fidelity, and licensing. The case remains a credible firsthand package finding, while the broader reusable agent-runtime interpretation is uncorroborated.
2026-09-02T04:23:14Z
The refreshed discussion remains repetitive criticism and speculation, with no independent package inspection or new evidence about how Codex invokes the bundled tools. Simon Willison’s firsthand finding remains credible, but the reusable agent-runtime interpretation is still uncorroborated.
2026-09-02T01:26:10Z
The refreshed comments add only speculative criticism about dependency size, fidelity and licensing, with no independent package inspection or new invocation details. Simon Willison’s firsthand finding remains credible, but the broader reusable-runtime interpretation is still uncorroborated.
2026-09-01T23:30:10Z
The refreshed discussion remains repetitive and speculative, adding no independent package inspection, invocation details, or evidence that the bundled tools form a reusable agent runtime. The case remains a credible firsthand implementation finding, but its broader architectural meaning is still uncorroborated.
2026-09-01T22:24:22Z
The refreshed comments remain speculative about provenance, licensing, bundle size and document fidelity; they add no independent inspection or invocation details. The case therefore remains a credible single-source implementation finding rather than a corroborated reusable-runtime pattern.
2026-09-01T21:47:51Z
The refreshed discussion adds scrutiny around bundle provenance, licensing and document fidelity but no independent confirmation or implementation detail. Simon Willison’s inspection remains a strong single firsthand line, so the runtime pattern is plausible but not yet corroborated.
2026-09-01T21:30:38Z
grounded: converges/medium — OpenAI’s reported bundling of LibreOffice is a consequential implementation of Scott’s “give the agent a workshop” and Code-First Architecture positions: system
2026-09-01T21:26:16Z
origin walked (codex/luna, conf 0.98): anchor hn.story.49527396 -> echo.blog.4060d5fb1d by Simon Willison
2026-09-01T21:24:48Z
case created — The reported packaging choice is a concrete and technically consequential implementation pattern with substantial early discussion.