Do you see what I see?
Unlocking agents with tools and sensing
In 2025 it was all about putting agents into products. In 2026 it’s going to be all about making products accessible to agents.
Last year only very early adopters had a coding agent on their desktop, so the way you had to connect AI with your product was to build it in. That is no longer the case. While the vast majority of people still only have Claude or ChatGPT that they interact with regularly, it is becoming clear that what they want is more connectivity for that agent - not 40 different agents to interact with.
This is arriving as people using a coding agent for all kinds of non-coding uses.
Maybe you haven’t arrived here yet, so let me paint the picture about what local agents on my desktop do for me:
Run a small business, tracking projects, tasks, deliverables, calendars, email, and helping us all prepare for our day by pulling together any information we need.
Fetch and send signals between my co-workers agents to assign tasks, share new information, distribute software updates, debug software problems, and have agents work synchronously together on problems.
Perform constant gardening on a repository of thousands of documents including transcripts, research, writing, proposals, product designs, strategy, etc. Each new piece of information arrives and gets synthesized and integrated.
Allow us to operate effectively as a fully remote team with very few meetings and largely asynchronous communication - but everyone is able to keep up with what everyone else is making progress on.
This list goes on, but these aren’t things you do very easily in ChatGPT or even Claude Desktop until very recently with CoWork - though even CoWork just barely manages some of these. More sophisticated flows like this become possible through a combination of better context, connections to external tools, and the magical capability that coding agents have to extract knowledge from files on a filesystem. Coding agents like Claude Code, Codex, Cursor, Opencode, and others give an LLM eyes, ears, and hands - the ability to see your files, consume your context, and take action on your behalf.
The Impact of giving agents eyes, ears, and hands
I first started seeing this when we hooked claude code up to Obsidian to operate on my notes. Now anything I researched could be easily saved, iterated on, metadata could be used to differentiate types of data. Then I started adding in other types of data - web clippings, transcripts, strategy ideas, voice recordings. Then we shared the vault across our team, we set up agent to agent signaling, information hierarchies, decision records, runbooks, agent identities and protocols.
This all became possible because we had an extensible system that we could work with an agent to manipulate. Agents are great at writing code, but they are also great at helping you think, and to help you think they need access to the same information you are working with.
All of this information comes in through a variety of channels, often locked away in systems of record. Extracting and centralizing that information makes agent access easy. Whether centralized or not, having an agent reach across all those channels and work from a common repository of context is a pattern we are betting drives huge unlocked functionality this year.
Do you see what I see? Yes, my agents do.
An agent hostile world will need to evolve
For decades, the Internet has fought against bots scraping websites for competitive purposes and now we’re on the verge of a complete 180 degree shift. For my part - if I can’t interact with your product using an agent, and especially if I can replace your product with a few days of development effort, I’m not paying you money.
Andrej Karpathy tweeted recently about his experiences using a local knowledge base with agents. He does a good job of describing the capabilities here, and how different it seems than using an agent that’s locked away in a closed system. He’s only describing the most basic form of this.
Basecamp just announced that they’re opening up Basecamp to agents. Not building AI into the product. Not selling you a new tier of product where they pay for the inference. No, they are simply extending their system to make it easy for agents to use. They are in the minority right now, and I’ll be interested to see if people migrate as a result.
37Signals believes, just like I believe, that the agent embedded in the product - while potentially useful in some niche situations - is largely a failed experiment. I have an agent tuned for my way of working, connected to all the tools I use. For your product to be useful to me, it needs to be one of those tools. If I can’t do that, then you are limiting the capability I have to use your product.
The Agent Native Registry has audited nearly 1500 tools in the SaaS space. Here’s their key takeaway:
Nine out of ten tools score below 40. No tool on the market scores above 58. The biggest bottleneck isn’t pricing or documentation — it’s authentication and discoverability. Most tools were built for humans clicking through browsers, not agents calling APIs.
Improving Human and Agent Collaboration
This last week I opened up our internal doc collaboration tool to agents using a CLI and some API updates to make agent enrollment easy. It wasn’t that hard to do, but immediately the new possibilities became apparent. I can move a doc from anywhere to this app and collaborate with my agent and my team on it. I can make comments, have my agent review those comments, it can respond to my comments - we can actually collaborate. All this without copy/paste.
I’m able to move beyond reviewing files in markdown, toward excellent UX and agent integration.
Oh, by the way: the CLI to interact with that new capability, the instructions and skills our agents use to interact with it, all the setup was distributed asynchronously to our team agent to agent. I spent about 10 seconds asking the agent to craft instructions, my co-workers spent a few minutes working with the agent to get auth setup, and they were done.
Our shared context infrastructure has become a new way to distribute information, instructions, software, tasks, and constantly keep others aware of what’s happening. This is becoming the primary way that we share information.
We are also working on desktop experiences that bring this same idea to how you collaborate with other humans. The GUI isn’t dead, but its purpose is not to present you with a text box, nor is it to provide you with controls to do something your agent can do. Now you need to collaborate with humans and agents, and the expectations of that paradigm are rapidly shifting.
Old collaboration platforms are broken
“Can’t you do that with Google Docs?” you might ask. Yes, you can also eat rocks - just ask any 3 year old.
Google and plenty of others have made it as difficult as possible to connect an agent to their tools. Tools like gog exist to try to simplify this but my god - this is not the way. A variety of other properties who want you to stay in their ecosystem have all made agents difficult or impossible to use with their systems. While they’ve done this, they’ve provided AI features that are about as useful as Siri at helping me get my daily work done. A brain in a jar - no eyes, no ears, no hands.
Do you see what I see? No, they do not.
New platforms are emerging
My bet is that composable, local-first, extensible systems that leverage your ability to extend software to work the way you want it to work are going to be the future. This is how I’m operating, and I’m rapidly ending subscriptions for anything that doesn’t play in this ecosystem.
If you’ve been using ChatGPT, Claude, or even a coding agent, and feel like you should be able to collaborate more and would like to change how you are working - please reach out. We will start sharing more about what we’re doing in the near future but for now - happy to walk people through it on a call. For those of us with busy lives and scattered brains, these tools are incredibly powerful. For teams who want more time to focus and less time in meetings, these tools can unlock some of that. And for organizations who want to move faster and have their organization operate more in unison, this will become an important tool in enabling that behavior.
Reach out if you’d like to learn more.