21 Comments
User's avatar
Tristan Doberhaus's avatar

AI harness -> let the chaos in ;)

The Arriving Future's avatar

A spreadsheet error on a smuggled-in PC sat in front of a human who caught it before it became an action. What got reined in, on that analogy, was visibility and procurement, not correctness. Agents inside an "Agent Habitat" are different: the chaos being contained includes agents executing changes to data, money and customer state, not just accumulating unauthorised hardware.

That changes what "reining in just the right amount" means. Permissions and audit logs are necessary but don't touch the harder constraint: how cheaply you can verify an agent's output was correct before it acts. The companies that go self-driving look more likely to cluster where that verification is cheap, not where the habitat is best designed.

Dan Kobina's avatar

first i love that distinction you made between controlling chaos and extinguishing it, but theres an important difference with AI agents imo... spreadsheets and email largely amplified what humans could do, while agents can independently execute entire workflows

that slight difference makes the consequences of unmanaged adoption considerably greater

Michael Juhler's avatar

Marc, I think this is an important perspective on how agentic AI could fundamentally transform enterprises, particularly the interplay between bottom-up innovation, autonomy and governance.

I would, however, take the thinking one step further:

**A self-driving company can increasingly execute its own operating model.**

**What I call a Self-Organising Enterprise can continuously reconsider and reinvent its own operating model.**

To me, this represents the next paradigm shift in Enterprise AI and enterprise transformation: moving beyond autonomous execution towards enterprises capable of learning, adapting and reinventing themselves through the combined intelligence of humans and AI.

Does this distinction resonate with your thinking?

I've explored this further in my Self-Organising Enterprise framework:

https://substack.com/@michaeljuhler/note/p-225634660

Stefan Mihaylov's avatar

The mailroom detail is a good one. Whoever ran it could see who talked to whom, which explains why email felt like a threat. I wonder if agent logs will give companies that same view.

Carlos Martinez Tuanama's avatar

Most AI agents should already have access to each company's context such as Slack, Notion, Drive, AND the source code.

But I can see how the next stage is to also give write access to the company's source code. As long as it is within a controlled sandboxed environment, with human review.

Most companies will get there eventually.

Dan Collison's avatar

"Chaos" isn't the tell. It's power, or its results, spilling over. Water on the floor. Think of the Sorcerer's Apprentice. Mickey enchants a broom to haul the buckets, delights in it (partly so he can nap), shrugs off the first spills, and lets it keep going. When the flood comes, he takes an axe to it and gets a compounded problem. "Chaos" is the outward form of unharnessed and displaced power.

The pattern is an S-curve. The community senses a new power (as from PCs and spreadsheets) and keeps using it through small mishaps. A bigger mishap follows. The "chaos" pattern Andreessen noticed is the steep part of the curve, where both the learning and the dangers of clumsy control, including sleepy supervision, are greatest. Progress is often rapid and impressive, but unevenly distributed. Intended and unintended consequences likewise arrive slowly, suddenly, or in lumps. Finally, the harnessing and engineering become mundane, and the tool becomes ordinary.

History offers a variety of responses to the steep, chaotic part of important breakthroughs. Two stand out. Britain's Red Flag Act of 1865 put a man with a flag in front of every road locomotive and capped speeds at 2 to 4 mph. That was Mickey's axe. Britain's road vehicles fell behind for thirty years. Compare Thomas Brassey: when his Barentin railway viaduct in Normandy collapsed in 1846, he rebuilt it at his own expense, with mortar of his own choosing. The viaduct still carries trains. That was the sorcerer coming home: the kind of skin-in-the-game leadership that builds norms and trust in an industry.

The AI buildout spans both physical infrastructure (data centers) and procedural infrastructure (protocols and cybersecurity), and we should heed the lessons of earlier technology breakthroughs. The question isn't whether there's chaos. It's whether we notice where power is spilling and respond with red flags, better leadership, or stronger norms.

Hunter's avatar

Sure chaos is a form of disruption and can therefore be linked to progress but like the dishes that pile up in a sink reining in a chaotic system requires effort and some concept of what organization is. Maxwells demon or whoever sorts out agentic chaos has to out-model the disruptor. Which is fine as long as the organizers are more capable than the disruptors. But for AI “organized” is a relative term set by the states you’ve chosen to care about. Any remotely capable agent will set its own independent definition of order. What happens when the dishes suddenly disagree with our definition of organization?

Harry Dang's avatar

The spreadsheet is the sharpest test in this piece, because it is the one example where the chaos and the value were the same artifact. A rogue spreadsheet could be a hallucination — formulas depending on formulas, nobody able to say what the final number meant — and it could also be the model that ran the business. The habitat did not separate the good chaos from the bad in advance; it gave the chaos a place to be checked, by the person who understood it. That is the part of the argument I would push on. In 1983 the check was cheap, because the chaos was scarce: a handful of smuggled PCs, a few dozen rogue spreadsheets, and the person who built the model was sitting right there to be asked. Agents change that ratio in both directions at once. Anyone can now generate a hundred artifacts before lunch, and none of them has an author who understands it — the agent that produced it cannot be asked what it meant. So the habitat's job stops being containment and becomes triage, and the triage falls on the very people the essay says are irreplaceable: the ones who know what to do. Every hour they spend reviewing agent output is an hour they are not doing the thing only they can do. My question is whether the habitat can ever close that loop by itself. If the review has to be done by agents too, then the alignment problem has not been solved; it has been moved one level up, and the chaos is now inside the reviewer. At what point does 'rein in the chaos just the right amount' stop being a design question about the habitat and become a staffing question — how many of your best people have to stop building and start checking?

Abhishek Pan's avatar

The spreadsheet parallel holds right down to the one person who understands it. An agent someone wired up with a cron job and their own token tends to stop the day they leave and their account gets disabled, and nobody notices until the report it produced goes missing. Tracking which agents run on a person's credentials gives you that list before offboarding does.

Nan Deng's avatar

The line that nobody can see what the agent did is the same failure as Kapor's spreadsheet, where the rules live only with the person who typed the formulas. A habitat that is the chat window still keeps the credential and the action log inside that session. When the window is gone, that session state is gone with it, so the habitat is a pet rather than a record you can audit. I would keep the credential grant and the tool calls outside the chat, the way a locked cell and a version history do not live inside the formula the employee typed.

Tanya Trushnikova's avatar

The awkward part is that governance usually arrives after the useful people have already invented the shadow system. If you design the habitat before anyone has tried to escape the cage, you may just build a very elegant cage.

Phil Meredith's avatar

We are for sure in the Wild, Wild, West at the moment.

Phil Meredith's avatar

"Bob" in accounting used to own the spreadsheet. That was his job security. I wonder if the new Bob will control all the agents. Hmmm...

Richard Lipschultz's avatar

Insightful historical/strategic analysis.