Cowork and the Way Things Are Going

We recently attended the Ormeau Labs X Base event and found ourselves sat next to Richard Nugent - founder and intellectual property expert who, like us, was using Claude regularly. Richard had gone further down the rabbit hole with skills, plugins, etc than a lot of developers. We had a good chat about a potential Windows OS release of Claude Cowork, which manifested when it dropped the next day.

Richard wasn't the only one at that event interested in Cowork. Ian Browne used it to publish a Founder Labs countdown timer just a day or two later. At the time we thought, "this is the way things are going."

For those unfamiliar: Cowork is Anthropic's AI workspace. Think of it as a desktop environment where Claude can work with your files, connect to your tools (Slack, Gmail, Notion, Stripe), and carry out multi-step tasks. Plugins extend what it can do into specific domains. Install a sales plugin and it researches prospects, preps for calls, and drafts outreach. Install a legal plugin and it reviews contracts. Anthropic came out the gate swinging with 10 or so official plugins.

These plugins were the horseman of the so called "SaaSpocalypse" - $285 billion wiped from SaaS valuations in three days. All because Anthropic shipped a bunch of markdown files.

Cowork plugins are, at their core, natural language content containing domain-specific knowledge. To give you an example: here's a small snippet from Anthropic's official marketing plugin showing how to optimise headlines:

Headline Formulas

How to [achieve result] [without common obstacle] — "How to Double Your Email Open Rates Without Sending More Emails"

[Number] [adjective] ways to [achieve result] — "7 Proven Ways to Reduce Customer Churn"

[topic]: what [audience] needs to know in [year] — "SEO: What Marketers Need to Know in 2025"

So if a free plugin can bottle up marketing strategy and Cowork can draft campaigns, segment audiences, analyse performance, etc then why are we paying random SaaS tools £15/seat/month?

The companies hit hardest all share a profile: subscription-based applications without a deep system of record. They are, at their core, interfaces on top of workflows. And when more of someone's workflow moves to AI workspaces like Claude, Copilot, and others - these products make less sense.

To Be or Not To Be “AI-Native”

AI-Native is a bastardised term. Companies slap a RAG chatbot on their product and call it "AI-Native."

Internally we talk a lot about the "Post Web" - the idea that the presentation-layer for the internet is dying out and most things will move towards intention-based economy, outcome-based pricing, and so on. That to us is what AI-native means. It's not about whether there's a traditional UI for the thing you're trying to build, though maybe that's an easy place to start thinking about it. It's about where you sit in the stack.

Here's an example. Our favourite feature on <ChatGPT/Claude/Deepseek etc> is Deep Research - it runs asynchronously until it's done, pulling together information, and hands you a report. But on all of these platforms, those reports are throwaway. They don't exist beyond the conversation session.

So we had this idea codenamed "Report Cloud" - a platform where research could build on itself. Old investigations could influence new ones, we could identify patterns across domains. Basically, a personalised Wikipedia.

We built V1 in a couple of days. Multi-agent research system and a platform to host the content. It totally worked.

But from a product perspective - it was all wrong. We were competing on the intelligence layer, trying to out-research Claude, OpenAI etc. You can't win that game when foundation models improve weekly. Worse, deep research already lives inside the AI workspace. We would be expecting people to leave their full-fat tooling to try our diet version.

So we ripped most of it out. Not unlike Homer Simpson trying to build a BBQ. We refactored it into a basic API accessible via Model Context Protocol (MCP) - in other words, something existing AI tools & agents could talk to.

Now we could run deep research in Claude where we do all our other work, and use the Report Cloud connector to publish results to our cloud. The MCP server self-describes its tools when Claude connects, so Claude can transform its research output into organised pages automatically.

"Couldn't you just use Notion for this?" Sure. We actually do use the Notion connector for other things. But Report Cloud can have whatever features we want - custom research templates, cross-investigation linking, etc. That's a broader topic about horizontal SaaS built for the masses versus something built by us for ourselves.

Anyway we're just using Report Cloud to illustrate the iteration from competing to extending. V1 tried to out-research the incumbent tools. V2 worked inside them.

We think that's the thought exercise for SaaS founders.

So What Do You Do?

The ground is still moving. But here's what we think today (18th February 2026):

These AI workspace tools are becoming (if not already) the default operating layer for knowledge work. That means standalone tools have to justify why they exist outside that layer.

Tara Simpson wrote something recently that stuck with us: "The hardest part of software development has never been writing code. It's understanding what to build and why."

If you own the data: Double down on being the system of record. Consider agents a target customer. Make your API the best way for agents to read and write that data. Think about how defensible the data itself is - could AI generate "good-enough" synthetic data for the same problem space?

If you own the domain expertise: Package it. Think about whether it's better as a standalone platform with its own UI or as something these workspace tools & agents can consume. Become the intelligence layer that foundation models can't replicate because it's specific to your industry.*

If you're infrastructure: Life is good when you sell picks and shovels.

* Right now the AI wars are in general capability - and especially coding. That's stabilising, after that providers will move up the stack into vertical domains

This is difficult to articulate in writing, but here goes: AI is changing how we create, which is changing how we think. As traditionally trained programmers, our work feels entirely different - but as consultants it's exactly the same. We help people solve problems. The matrix of how much help that person needs and how we provide that help feels like it's undergoing a big shift because of this technology, but honestly it was always different anyway, project to project, client to client.

If all that's left is helping people think, well then we're here for it.