Founder Mode in Era 3: When Half Your Team Can Only Read

On October 6, 2024, Tom Blomfield published Founder Mode. About a thousand words arguing that founders should obsess over the details their managers want to hide. Eighteen months later the essay still travels. What it does not yet name is the new mechanism.

Founder mode in 2026 is no longer just reading the meeting notes. It is reading the fleet's output and writing the missing context the fleet cannot generate for itself. The scarce role in the building is whichever operator does that loop fastest.

I have been watching this play out at Well. We are 12 humans and roughly 28 specialised AI agents, each one shadowing a role on the team rather than a person. We ship 35 to 50 features a month. The bottleneck is not engineering output. The bottleneck is the manager who reads the fleet's output, spots which agent is operating on stale context, and writes the missing standard before the next loop runs.

Founder mode is the right name. The mechanism changed.

Tom's essay describes a stance. The founder who refuses to be managed away from the details. The CEO who reads every customer support ticket at 2am. The kind of leader who lands in every PR thread before merge.

That stance had a hidden mechanism in 2024. The founder absorbed context from a hundred small conversations. Standups, hallway questions, Slack DMs, customer calls. Then re-injected the context through a thousand small decisions. The mechanism worked because every team member was a human who could receive context the same way it was absorbed.

By 2026 the stance has not changed. The mechanism has. The team is no longer all-human. Half the team reads, does not listen.

The agents on your fleet cannot attend standups. They cannot rewatch Looms. They cannot follow a Slack DM thread. They were spawned five minutes ago with zero context, and they will execute against whatever was written down. So the founder-mode stance still works. The mechanism is now writing.

The audience changed

The clearest statement of this came from Andrej Karpathy at Sequoia's AI Ascent 2026. He called modern AI agents ghosts shaped by documents. The shape is the corpus. The shape is everything the agent will ever know about your business.

That is the new audience. It is not new because it is hostile. It is new because it cannot ask follow-ups. It cannot read between the lines. It cannot reconstruct your reasoning from your tone. It can only read what you wrote down.

This is why every era of company-building eventually rediscovers writing. Amazon discovered it under Bezos in the early 2000s. PowerPoint banned by email on June 9, 2004. Six-page narrative memos. Silent reading rituals at the start of every meeting. Stripe ran the same playbook a decade later. Patrick Collison sending emails with footnotes. Strategy memos read like research papers. Alan ran it in French. 1,560 GitHub Issues at a 50-person company by September 2018, all visible to every employee.

The startup decade between roughly 2010 and 2022 ran the opposite stance. Verbal-first. Two-pizza teams. Loom-instead-of-docs. That stance was correct for its audience. Humans in a shared room. Humans in a shared Slack channel. Then the audience changed. The two-pizza team now includes the model.

Where the bottleneck sits now

I am going to name the role that catches this drift first. It is not the CEO alone. It is whichever person on the team reads the fleet's output most carefully and writes the missing context fastest.

At Well, that role rotates by domain. For engineering, it is whoever has been running the build longest and noticed which prompts the agents are operating with. For GTM, it is whoever has watched the most lead enrichment outputs and spotted which counterparty signals the agents are missing. For design, it is whoever maintains the Storybook canvas and writes the standards the design agent reads against.

What every one of those people has in common is this. They read agent outputs the way Patrick Collison read prose. Closely. Suspicious of every conclusion. Looking for the gap between what the agent said and what the actual state of the world is. When they find a gap, they do not coach the agent. They write the missing standard.

This is founder mode for an Era 3 team. The manager is not a project administrator anymore. The manager is the writer the fleet was missing.

What it looks like in practice

We ran an agent context audit at Well on April 7. Thirty days of fleet output reviewed. Four critical fixes surfaced.

1. The watcher prompt was carrying 30 to 80k tokens of irrelevant context into every spawn. Every agent was reading material that had nothing to do with its task. The fix was a per-agent essentials strip.
2. Branch-prefix routing rules were missing from the domain map. Agents could not tell which domain owned which branch. The fix was a written domain-prefix table.
3. The allowed_tools field was not enforced at spawn. Some agents could call tools they should not have access to. The fix was an enforced loadout.
4. The standup workflow was implicitly coupled to one agent that did not need to be on its critical path. The fix was a decoupling, written down as a tier-0 pipeline standard.

Every one of those fixes was a piece of writing. None of them changed how the agents work. They changed what the agents read.

The hours that produced those fixes went into reading agent outputs and writing standards. Not into agent training. Not into prompt engineering. Into the kind of work Tom describes in Founder Mode. The founder who reads every detail. Except now the details include the rendered prompts the agents loaded with, and the standards live in the repo where the next spawn will find them.

What I got wrong

I used to think the scarce thing on an AI-native team was the AI itself. The agents. The models. The prompts. The frontier-facing parts.

It is not. The frontier moves fast. The scarce thing is the human who can read what the frontier produced and write what was missing. The agent picks up the standard on the next spawn. If the standard is not there, the agent re-makes the same mistake on the next 50 spawns.

I also thought writing was a documentation problem. A library to maintain. A wiki to keep clean. That framing is wrong. Writing is not the library. Writing is the closing of the loop between what the agent did and what the agent should have done. It is operational, not archival.

The corollary is uncomfortable. A team that cannot write its standards down cannot grow a fleet. Not because the fleet costs too much. Because the fleet has no context to inherit.

What is still scarce

Features were never the value. Code was never the moat. In an Age of Abundance, the cost of building those features collapsed by roughly 30×. Most of the work the engineering team used to do is now free. I wrote about this last week in the SaaS Idiot Index piece.

What stayed scarce is the closing of the loop. The reading of the fleet output. The writing of the standard the next agent will inherit. The kind of management work that was easy to neglect when every team member was a human who could hold context in their head, and is now load-bearing because half the team has no memory between sessions.

This is what YC's 2026 Request for Startups calls the company brain. A system that pulls knowledge out of every fragmented source, structures it, keeps it current, and turns it into an executable skills file for AI.

That ask is not about a Notion-flavored database. It is about whichever operator in the building is doing the founder-mode read-and-write loop. The company brain is what is left over when that loop runs for long enough.

Closing

At Well, we left Figma because design lives in the codebase. We are leaving Notion because specs live in the repo and GitHub issues. The agents read the repo. Fragmenting context across surfaces just blinds the fleet.

But the move that compounded the most was hiring for the reading-and-writing loop. The managers who catch the drift. The ones who spot which agent is operating on stale context and write the standard before the next loop runs.

Founder mode never went away. The audience changed. The scarce role in a 2026 team is the human who can read the ghosts and write what they need to see.

Maxime Champoux is the co-founder and CEO of Well. Tom Blomfield's Founder Mode essay is here. His Age of Abundance essay is here. The full research consolidation on the documentation pendulum is in article #54 — Writing Culture Is Back. But Not For Us.. YC's 2026 company-brain ask is here.