Nothing Was Built for Working (With AI Agents)

Count the tools your team touched today. A chat app for talking. A doc editor for writing. A drive for storing. A project tracker for tracking. A wiki for remembering, in theory. Each one is genuinely good at its job, and together they've defined what "working" looks like for fifteen years: work is the thing that happens in the cracks between them.

You've felt the cracks even if you've never named them. A decision gets made in a thread on Tuesday and by Thursday nobody can find it. The doc says one thing, the tracker says another, and the truth is in a DM between two people, one of whom is on vacation. Someone spends forty minutes writing a status update that summarizes information that already exists in four other places. Every handoff between tools is a small leak, and a team of any size leaks all day long.

We mostly stopped noticing, the way you stop noticing a squeaky door. The tools were built for a specific purpose each, and the purpose was never "the work itself." Chat was built for conversation, and it's excellent at conversation, which is why decisions drown in it. Docs were built for the artifact, and they're excellent at artifacts, which is why nobody can tell which artifact is current. Trackers were built for the map, which is why the map and the territory diverge by Wednesday. The actual substance of work, the context and decisions and handoffs and half-finished states, has no home. It lives in the seams, and humans became the couriers who carry it across.

Here's the thing that turned this from an annoyance into a real problem: the couriers were the tolerable cost when everyone doing the work was human. That stopped being true this year.

AI agents are joining teams. Not as a feature bolted onto one tool, but as participants that take on real chunks of work. And the first thing every team discovers when they try to put agents to work is that the agent has the same problem the new hire has, only worse: it needs context, and the context is scattered across five tools and three people's heads. A human new hire spends a quarter absorbing that context through osmosis, meetings, and embarrassment. An agent doesn't get a quarter. It gets whatever you paste into the window.

So teams do what teams always do. They improvise. They paste the relevant thread, attach the doc, summarize the history, and hand it all over with the task. It works, sort of, the way carrying water in your hands works. Then the agent finishes and the output has to be carried back, posted somewhere, explained, reconciled with whatever changed while it was working. The human is now a courier for the agent too. We automated the work and kept the leaks.

It gets more interesting when there's more than one agent. Real work is rarely one task. It's chains: research feeds a draft, the draft needs checking, the check raises a question only a person can answer, the answer changes the draft. When agents can hand work to each other, and tag a human when they hit something above their pay grade, and pick the work back up after the human answers, you start to see what a genuinely mixed team looks like. Also what it demands. Chains like that need a shared place where the state of the work actually lives. Not five places. One. A place where a person can see what the agents did while she was in a meeting, where an agent can see what the humans decided while it was idle, where "what's the latest" is a question with exactly one answer.

None of our current tools is that place, and it isn't their fault. They were designed around a quiet assumption so universal nobody ever wrote it down: every participant is a person. Chat assumes readers who skim and infer. Docs assume authors who hold context in their heads between sessions. Trackers assume someone will remember to update the ticket, which was optimistic even for humans. Add participants who work at machine speed, around the clock, with no ability to lean over and ask what a cryptic comment meant, and the assumption doesn't bend. It breaks.

There's a precedent for this moment, and it's instructive. Every prior generation of work tools was born from one question changing. "How do we talk across distance?" gave us email and then chat. "How do we write together?" gave us collaborative docs, and it seemed like magic in 2006. Ea

ch time, the winning products weren't better versions of the old category. They were answers to the new question. The new question is already here, being asked in slightly different words inside thousands of companies: how does a team work when some of the team isn't human?

I don't think the answer is a better chat app with an AI summarizer, or a doc editor with a copilot in the margin. Those are the old categories wearing a new hat, and the seams between them are precisely where the problem lives. The answer, whatever shape it finally takes, has to treat the work itself as the first-class citizen. The context, the decisions, the handoffs, the current state of the thing being made, held in one place that humans and agents inhabit together, on equal footing as participants rather than as users and features.

That's a category that doesn't quite exist yet. Pieces of it are appearing at the edges of existing products, and anyone paying attention can feel the pull in one direction. My honest bet is that within a few years, "where does the work live" will be a question with an obvious answer, and we'll look

back at the courier era, the pasting and the reconciling and the forty-minute status updates, the way we now look back at emailing documents named final_v7_FINAL.

The teams that adapt earliest won't just be faster. They'll be organized differently, around a different center of gravity. The tools we have were built for talking and writing and storing, and they did their jobs well. Nothing was built for working. Someone is going to build it, and the moment it exists, the cracks we all stopped noticing are going to become impossible to unsee.