Your company brain needs a librarian

Ah, the company brain. Who amongst us wouldn't want a company brain? The promise is enticing: all of your collective knowledge, shared and available on demand, so everyone can tap into the peak wisdom the company has to offer and make the most informed decision every time. It not only sounds intriguing, it also sounds like the kind of problem AI should be solving for us. And there's no shortage of products and social media posts floating around that claim to have done just that. We're still skeptical.

To jump ahead and steal from our first article this week, which argues that the name "brain" is itself a big part of the problem:

A company context repository is a less exciting name for sure, but it describes what it actually is.

>

It also leaves room for the rest of the architecture, which many of us are actively building. The harder systems are the ones that govern how that context gets created, selected, validated, updated, and propagated through work, and that problem is substantial enough without pretending the context itself is doing the thinking.

As that quote suggests, a company "brain" is not one thing. It is a series of capabilities layered together to, ideally, create something broadly and deeply useful. But some of those capabilities are more difficult to realize than others. Here are a few choice examples.

Connecting to data sources

This part is relatively well solved. Thanks to MCPs (as well as other all-time great 3-letter acronyms like CLIs and APIs) getting data out of systems and into agents' grubby little hands is generally not the problem.

Governing access

But you know what is the problem? Making sure the right people have access to the right data. Lock the data down too much and nobody can do anything useful, nullifying the benefits for everyone. Give broadly accessible agents access to certain kinds of data, and you might be in trouble (especially in regulated industries). So you need a system to govern what data an agent has access to depending on the environment or the user who requests it. Today this is most commonly solved with an MCP Proxy or Gateway. Compared to connecting sources, this is a harder problem in that it requires more work, but it's a relatively solved or solvable problem.

Curating context

So you've put in the work. You've connected the data, and you're governing the data properly. Fantastic. Now which data is the right data? How many companies have 3 different versions of the roadmap in Confluence, and everyone who needs to know knows which is the most up to date one, the one to look at? Unfortunately agents don't know without a hint or a pointer. Access to all of the data is, at best, a crapshoot if the data isn't curated. This unfortunately is not a problem with a crisp technical solution. Certainly someone (or something) can do the curation. If done well, and done consistently, this is a great solution. At any meaningful scale this is really hard to pull off, and that's where we see most company brains break down.

So where does this leave us? The building blocks still matter, but the magic button doesn't exist yet. We believe that investing in data access, data governance, and data quality are all extremely worthwhile. But you need to equip your team with the right expectations, the right hard skills to be able to use the tooling effectively, and the right guardrails to identify when things go off track. The "company context repository" doesn't sound as glamorous or smart as a "brain", and we agree that maybe that's a good thing.

Here's what we're reading this week.

Stop Calling the Company Brain a Brain

Josh Rosen. We already spoiled this one in the intro but it's worth a full read. Josh lays out a compelling case for why the amorphous thing we've all taken to calling a company brain isn't a brain at all.

What we've learned scaling AI coding agents at Spotify

Spotify released Xirp in free public beta, their "agentic development environment" (it's a harness that wraps other coding harnesses; you figure out what to call it). Naming choices aside (is it a homophone of ZIRP?), it's an interesting piece for a number of reasons.

First, we've now talked to or heard of several teams that are building "harness-agnostic" (or as Spotify describes it "vendor-neutral") wrappers that allow easy switching between the likes of Claude Code, Codex, Gemini, etc, sometimes within the same task. It protects against lock-in, gives developers freedom to use the tools that work best for them, and for the most advanced users it's like model switching on steroids. And perhaps most importantly it allows building user- and developer-experience improvements on top of the abstraction, so you can do things like share, resume, or collaborate on sessions.

When a new model emerges, our engineers transition without reconstructing their environment. They are able to move fluidly between coding agents, preserve context across sessions, and orchestrate parallel workstreams.

Second, there's a whole section on "the multiplier effect" of connecting Xirp to Portal, which is an engineering-specific company brain of sorts. The Portal landing page describes its capabilities this way:

Catalogue your systems, govern ownership, and manage the context needed to build with confidence.

Sound familiar?

The top reasons AI stalls inside companies

Claire telling hard truths on X this week.

The blocker is never tools or intelligence. Human systems, human problems!

How my vibe coded website looks like a designer made it

We just think this one is fun. Go read it, it's full of interesting tips for iterating on design with agents as a non-designer. But to spoil the answer to the question "How does one make a vibe-coded website look like a designer made it", the answer seems to be: care a lot about the output, iterate obsessively, and put in lots and lots of effort until you're happy. So, vibe-coded perhaps but not what we would consider "vibe-designed". A great example of how to leverage AI to raise your ambition, but also a sharp reminder that the only true antidote to AI slop is care and effort.

One thing to try this week

xAI launched Grok Bot, their desktop app for creating a team of AI agents to do your bidding. To set expectations it's better to think of it as a Cursor product than an xAI product (SpaceX, which merged with xAI in February, agreed to buy Cursor's parent company Anysphere in June). Which is to say, we've been pretty impressed so far. It's in early access, but if you get a chance it's well worth a spin.

And if you can't get access to Grok Bot, don't despair. We have another thing to try that's available to anyone: Dex Horthy showed off his /show-me skill this week, which encourages agents to explain content through diagrams and code snippets, not prose. Especially useful if you're still befuddled or enraged by how the latest Anthropic models communicate.

See you next week!

— Claire + Zach