What Will Conway See When It Watches Your Team Work?

Last week, Anthropic accidentally published half a million lines of Claude Code source to a public registry. Buried in that leak was something called Conway, an always-on agent environment that nobody at Anthropic has announced. Nate B. Jones did an excellent breakdown of what the leak reveals, and I want to build on his work because I think the real story isn't the technology. The real story is what this technology forces every organization to confront about itself.

But before we get to Conway, let's talk about what happened yesterday.

Mythos Just Ended the "We Have Time" Conversation

Anthropic launched Project Glasswing this week, giving a small group of partners access to Claude Mythos Preview, a model so capable at finding security vulnerabilities that Anthropic decided it was too dangerous to release publicly. In just a few weeks of testing, Mythos discovered thousands of zero-day vulnerabilities across every major operating system and web browser. The oldest was a 27-year-old bug in OpenBSD, an operating system whose entire reputation is built on security. Decades of professional security audits, billions in cybersecurity spending, and an army of talented researchers missed what this model found in weeks.

Here's the part that should keep you up at night. Anthropic didn't train Mythos to do this. These capabilities emerged as a side effect of general improvements in reasoning and code. The model just got smart enough that breaking into things became something it could do on its own. The total cost to scan OpenBSD across a thousand runs was under twenty thousand dollars. The specific run that found the critical bug cost less than fifty bucks.

The security industry didn't get a warning. It didn't get a transition period. It got a phone call that said hey, you know that thing you've been doing for thirty years? An AI just did it better, faster, and cheaper, and by the way it found a few thousand problems you didn't know you had. Good luck.

Now hold that feeling. Because Conway is that same disruption pointed at your org chart.

What Conway Actually Is

According to the leaked source, Conway operates as a standalone sidebar inside the Claude interface. It's not a chat window. It's an entire agentic environment with its own extension format, browser control, tool connections, and the ability to be woken up by outside events while you sleep.

Nate B. Jones painted a picture of what a Tuesday morning looks like six months after you set this up. Conway has been running overnight. It noticed three emails that matter to you, not because you wrote rules but because after months of watching you work it learned which ones matter. It drafted responses to the easy ones. It flagged the one from your VP but didn't touch it because it knows you need to see that yourself. It checked Slack, drafted a reply to a technical question using context from a design doc you reviewed last month, and pulled the latest numbers for your board prep because your calendar says that meeting is at ten.

You haven't typed a word yet.

About a third of what Conway did overnight might be wrong. The email draft might misread tone. The Slack reply might be technically off. But the net is still positive because the speed of iteration compensates for the imperfection. This is important. The value isn't accuracy. The value is velocity combined with human judgment at the review layer.

The Stories Leaders Tell Themselves

Every organization that matters will face a version of Conway within the next twelve to eighteen months. Not necessarily from Anthropic. Google and OpenAI are converging on the same architecture. The persistent agent layer is coming regardless of which logo is on it.

And when it arrives, it will force a decision that most leaders are not prepared to make. So let's talk about the stories they'll tell themselves to avoid making it.

"AI is still experimental. We have time." This is the story the security industry was telling itself last month. Then Mythos found vulnerabilities in every major operating system and browser that decades of professional auditing missed. The experimental phase ended and nobody sent a memo.

"Our industry is different." Too regulated. Too complex. Too relationship-driven. This is the same story every industry told about the internet in 1998 and about cloud computing in 2010. The specifics change. The denial doesn't.

"Our people are our advantage." This one is actually true, but not the way they mean it. They mean smart people make the technology unnecessary. The reality is that smart people are exactly who you need to identify and build around before the persistent layer arrives. Your people are your advantage only if you know which ones are actually driving value. Most organizations don't.

"We'll be fast followers." This works when the advantage is a product you can buy and install. It doesn't work when the advantage is accumulated behavioral context over time. The first mover doesn't win because they have better technology. They win because they have six months of compounding that the late adopter can never shortcut. You cannot fast-follow institutional memory.

"We tried AI and it didn't work." They bolted a chatbot onto a broken process. The chatbot performed like a chatbot bolted onto a broken process. And now that's the reference point for every AI decision going forward. This is like test-driving a Ferrari in a parking garage and concluding cars are overrated.

The Fork in the Road

Here's where it gets uncomfortable. A persistent agent like Conway doesn't care about your org chart. It cares about your commitment network, the actual pattern of who requests what from whom, who delivers, and who drops the ball. Every organization runs on these invisible networks of promises, requests, and assessments flowing between people. The hierarchy is just scaffolding that tries to make those commitments predictable.

Conway will learn the real network, not the official one. And this creates exactly two options.

Option one: you figure out where value actually lives in your organization before you deploy the persistent layer. You identify the people who consistently originate commitments that produce results. The ones who can walk into a messy situation with ambiguous goals and competing priorities and produce clarity about who needs to do what by when. You build the agentic architecture around those people.

Option two: you skip that step, deploy Conway on top of the existing structure, and the persistent agent faithfully learns and accelerates whatever is actually happening. Every vague commitment gets automated. Every diffusion of responsibility gets encoded. Every hallway workaround gets baked into the agent's behavioral model as if it were best practice.

For a few months option two looks like it's working because everything is moving faster. But faster isn't better when the underlying network is broken. It's just faster toward the wall.

And here's why option two is more dangerous than any previous technology failure. When Lotus Notes was deployed on a dysfunctional organization, it just sat there underutilized. Expensive shelfware. Embarrassing but survivable. When a persistent agent is deployed on a dysfunctional organization, it doesn't sit there. It actively compounds the dysfunction. It makes promises nobody authorized. It triages based on political patterns instead of actual priority. It learns that the real decision happens after the meeting and starts routing around the official process just like everyone else does.

If you remember the Ethereum DAO, you know how this ends. The DAO was the experiment in removing human judgment from organizational governance. Encode the rules, let the code run the show. It lasted about three months before someone exploited the gap between what the rules said and what the rules meant. The code did exactly what the code said. The problem was that the code couldn't handle situations its authors didn't anticipate. Conway deployed on a broken commitment network is the DAO at enterprise scale.

Finding the Players

So how do you actually run option one? How do you identify where value lives before you hand the keys to a persistent agent?

You follow the commitments.

Pick someone in your organization. Ask them a simple question: when you need something done that actually matters, who do you make the request to? Not who the org chart says you should go to. Who do you actually go to?

Then go to that person. Ask the same question. Follow the chain.

At some point the chain loops back or converges. Multiple starting points funnel to the same node. That's your player. Not because anyone declared them important but because the live commitment network routes through them organically. The organization already knows who these people are. It just hasn't said it out loud.

You can sharpen this by running a commitment audit on real projects, past and present. Take something that had actual stakes and trace the commitment chain backward from the outcome. What was the commitment that produced this result? Who made it? Who made the request that triggered it? Who saw the situation clearly enough to know what needed to be asked for? Do this across five or six projects and the same names surface.

The people whose names show up at the junctions where vague intention became specific commitment, where stalled progress got repaired, where assessments turned out to be accurate, those are the nodes you build around.

If no consistent names emerge, that's your answer too. You don't have the players on staff. Hire from outside. And not as a blame sink for when it goes wrong.

The Compensation Question Nobody Is Asking

There's a deeper issue here that Conway forces into the open. The traditional employment deal is straightforward: you show up, you do work, the company pays you, and when you leave you take your skills and reputation with you. The company keeps the output.

A persistent agent breaks that deal. Because now the company isn't just capturing your output. It's capturing the pattern of how you produce output. Your judgment. Your triage instincts. Your ability to read a room or an inbox and know what matters. Once that's embedded in the agent's behavioral model, the company has a version of your professional intuition that keeps generating value after you walk out the door.

If the persistent layer is capturing value that has a tail, the compensation should have a tail. A bonus doesn't match the shape of what's being extracted. Equity with post-departure vesting is a better structural fit. It says your behavioral contribution to this persistent layer is an asset that appreciates over time, and you should participate in that appreciation.

But let's be honest about the prerequisites. A lot of this depends on whether what the agent learns is actually accurate and valuable enough to matter. If the behavioral model is only sixty-five percent right after six months, the lock-in argument weakens. You're not leaving behind a copy of your professional judgment. You're leaving behind a rough sketch that still needs heavy supervision. These are empirical questions that haven't been tested in the real world yet.

The Question That Matters

David Graeber argued in Bullshit Jobs that a significant percentage of roles exist primarily to justify their own existence. The organizational structure needs the headcount for political reasons and nobody can measure the gap between activity and value creation with any precision.

A persistent agent layer makes that gap visible. Not because it replaces people but because it creates an x-ray of who actually drives outcomes versus who just touches the ball on the way through. The political cover for roles that exist without producing proportional value is about to get very thin.

Which means the organizations that thrive in the Conway era will look radically different from the ones we have now. Smaller. Flatter. Built around a small number of people who can structure and own commitments, amplified by agents that extend their reach. Not because technology replaced people but because technology made it impossible to pretend that every layer of the org chart was earning its keep.

This is coming. Mythos proved the timeline is shorter than anyone expected. Conway or something like it will prove that the same acceleration applies to organizational structure. The only question is whether you do the hard work of understanding your real commitment network before the persistent layer arrives, or whether you let the agent learn from whatever it finds.

One of those paths leads somewhere productive. The other one is a toddler with a chainsaw.

Choose carefully.