We've been saying knowledge work trails the coding world by about 18 months. In October, in Four Years, Then Four Months, we wrote: "Knowledge work is about 18 months behind this revolution, and for good reason."
The path coding opened.
Coding led the way. For a stretch, developers copied code into a chat window and copied the answers back. In 2025, the tools changed. Harnesses like Claude Code put AI in the codebase itself, where it could read the files, write to them, and stay in the work. The conversation moved from the model alone to the model plus the harness.
Knowledge work started down the same path, later, and for a simple reason. People who were not developers began using tools like Claude Code and Codex as a way to do knowledge work on their own computers. Anthropic and OpenAI both picked up that signal and built products that were easier to use than a terminal, so the same kind of harness could sit in ordinary folders and files. Anthropic released a research preview of Claude Cowork in January 2026 and made it generally available in April. OpenAI shipped ChatGPT Work along the same lines. In both cases, the idea was to bring AI to the place where knowledge work already lives, instead of asking someone to paste a piece of that work into a prompt.
We already see knowledge work benefiting from AI in the field. People are using these tools on real documents and real workflows, and the work is getting better. The pattern we described is showing up in the products, and in how people use them. It is still early.
What the roadshow shows.
Anthropic priced its initial public offering on November 18, with shares set to begin trading on the Nasdaq on November 19 under the symbol ANT. In the roadshow video at ipo.anthropic.com, they included a slide titled "Making fast progress across knowledge work."

Image: Anthropic's IPO roadshow video, ipo.anthropic.com.
In their telling, coding crossed a threshold in November 2025 and cybersecurity crossed one in April 2026. On the coding curve, earlier marks show autocomplete, then Copilot doing one task, one file at a time, before that November threshold. Knowledge work is at early agents today, with the threshold ahead, and then autonomous fleets of agents. A footnote on the slide says the curve is illustrative, and that a threshold is "the release after which long-running, multi-agent autonomous work became the frontier default." They do not give a date for when knowledge work crosses that line.
They are expecting a step change for knowledge work like the ones the slide shows for coding and cybersecurity, and that is exciting.
How knowledge work gets there.
We can understand how code and cybersecurity got there, given the nature of that work. Code is a relatively uniform artifact. You can point a system at a repository and let it work. We're much more curious how knowledge work gets there.
We understand why it is taking longer. Knowledge work is fundamentally different. It has structure, but it is less uniform than code. It lives in documents, spreadsheets, decks, email threads, meeting notes, and the workflows that connect them. That is a harder environment to take through a threshold of long-running, multi-agent work.
If any company can push knowledge work through that step change, it is Anthropic.
Sources
- Brandon Gadoci, Four Years, Then Four Months (October 22, 2026)
- Anthropic, Anthropic prices initial public offering (November 18, 2026)
- Anthropic, IPO roadshow, including the slide "Making fast progress across knowledge work"
- Anthropic release notes for Claude Cowork, as cited in Four Years, Then Four Months (research preview January 12, 2026; general availability April 9, 2026)
- OpenAI's launch of ChatGPT Work, as cited in Four Years, Then Four Months (July 9, 2026)
