Every Company Is Now a Mini-AGI

Jack Dorsey, on Sequoia's podcast in April 2026, dropped a line that's been turning over in my head: every company can now be a mini-AGI. He was pitching Block's internal vision. The phrase sounds like a deck slogan. From inside the work I've actually been doing with agentic engineering since October 2025, it stops feeling like a slogan and starts feeling more or less right.

I want to take Dorsey's frame seriously, not as a forecast about the future, but as a description of what's available right now if you're willing to organize your company around it. Because most companies aren't, and the ones that do will look very different from the ones that don't.

What Dorsey is actually claiming

The frame goes something like this: a company is a thing that takes inputs (signals from the market, capital, people), does cognition on them (analysis, decisions, prioritization), and produces outputs (products, services, allocations). For most of the history of companies, the cognition was distributed across a lot of human heads, coordinated via meetings, documents, and middle management. Slow. Lossy. Hard to scale.

What's different now is that a lot of the cognition can be done by AI systems running as agents. Each one a specialist. Each one able to read a brief, do work, return output, get reviewed, and iterate. Wired together, they're a cognitive layer the company runs on, not a productivity tool individuals reach for.

That's the part that makes "mini-AGI" not a stretch. A network of agents that observes, reasons, and acts on behalf of a company is, for that company's purposes, a general intelligence. It doesn't need to solve poetry or theoretical physics. It needs to read the customer support queue, the GitHub issues, the metrics dashboard, and the product spec, and produce useful work. That's a much narrower bar than AGI in the academic sense, and it's a bar that today's models clear in many domains.

What this looks like in practice

This isn't theoretical. Run parallel agents against tight briefs — Claude Code, Codex, Gemini CLI, Snowflake Coco — with several work streams active at once, and a pattern shows up at the company level, regardless of headcount.

The org chart flattens. Middle layers have less to do. Work moves between briefs, parallel work streams, and reviews; the human in the loop is the one whose taste decides what ships.

Headcount stops being the lever. When the bottleneck shifts from lines-of-code throughput to brief quality and review judgment, more bodies don't help. The lever is orchestrators, not headcount.

Roles compress. "Backend." "Frontend." "DevOps." Categories from a world where each was a person's full time. Now they're skills any one person picks up because the agent does the typing. The remaining specialty isn't a stack — it's taste and judgment about systems as a whole.

Cycle time shortens. The time from "we should try X" to "X is shipped to a small set of users" compresses. That changes what kinds of bets become rational. Things that were too speculative to staff become cheap to test.

That's the local version of the dynamic Dorsey is pointing at. The shape of the work changes; the size of the org stops predicting its output the way it used to.

What it doesn't change

Three things stay exactly as they were.

Distribution still wins. A company that ships ten products with no users is still a company with no users. The cognitive layer accelerates building. It does not accelerate distribution. Marketing, sales, community, partnerships — those are the same human-coordinated work they always were. The companies that will benefit most from agentic engineering are the ones that already have distribution and need their build velocity to catch up.

Taste still matters. When you can produce three variants of a feature in an afternoon, the question isn't can we build this, it's which version is right. Without strong opinions about what good is, you ship variations of mediocrity. Taste isn't a soft skill anymore; it's the binding constraint.

Trust still has to be earned. The chain-of-thought log every agent produces is auditable. That's necessary but not sufficient. Customers, regulators, employees — all of them still want a human accountable for outcomes. "The AI did it" isn't a reason. The orchestrator owns the output.

What this means for the people in companies

If Dorsey's frame is right, the people who thrive in mini-AGI companies are different from the people who thrive in traditional ones.

They're the people who can hold a system in their head — eight things shipping in parallel, each at a different stage, each with its own brief and eval — and make the right call when one of them goes sideways. They're the people who can write a prompt that's also a contract. They're the people whose taste is sharp enough to pick the right variant out of three without agonizing.

That's a different hiring profile. It's also a different career path. The traditional progression — junior → senior → staff → principal — was built around increasing implementation skill. The new progression is around increasing orchestration skill, which has different feedback loops and different signals of mastery.

I don't think most engineering organizations have figured this out yet. The ones that do will pull ahead. The ones that don't will be confused for a few years about why their hiring isn't producing the velocity it used to.

My pragmatic take

Dorsey's framing is grand. The day-to-day version is more boring and more useful: agentic engineering changes the company shape, not the engineer's day. If you're a CTO, that's the planning question to wrestle with. Not "should our engineers use Claude Code." That's table stakes. The real question is whether your org structure, your hiring funnel, your decision-making cadence, and your evaluation rubrics are ready for a world in which the implementation step takes minutes and the judgment steps take hours.

If they're not, that's where the work is. And it's not technical work. It's organizational work, which is the part most engineering teams are least excited about and most under-resourced for.

The mini-AGI is here. Whether your company turns into one is up to leadership, not the model providers.