---
title: 'The start of the Super Human era'
date: '2026-05-19'
description: "Agents aren't the future. Humans who think with models are. Here's why they're already pulling ahead."
tags: ['agents', 'AI', 'agentic', 'humans', 'software']
published: true
---

Fully autonomous agents aren't the future. Humans who think with models are. And
they are already pulling ahead of everyone who's still waiting for models to
mature.

---

## What is a Super Human?

A Super Human is not someone with better tools. It's someone whose bottleneck
has shifted.

Information access has been the biggest bottleneck to figuring things out
throughout human history.

Models do in seconds what used to take weeks of reading.

So the bottleneck shifts to judgment. Reasoning, decision-making, the ideas
you can't quite explain. The more context you have, the better the decisions
you make. Models take care of the context while you develop the right
judgment.

In practice this shift looks small in the moment and huge in aggregate. Looking
something up takes minutes instead of an afternoon. Drafting takes one pass
instead of three. The hard part of the day stops being *finding the answer* and
becomes *deciding which of three answers is the right one to act on*.

Judgment isn't built from past information alone, or from what's in front of
you right now. It comes from past experimentation, from things you noticed
without trying to. Models don't have that.

Demis Hassabis made this point about Go: top players often choose moves they
can't fully explain. They just feel right. [DeepMind did teach AlphaGo that
intuition](https://www.technologyreview.com/2022/02/23/1045016/ai-deepmind-demis-hassabis-alphafold/),
but only Go's. Yours is still yours to build.

So we run a loop. Explore with models, build judgment from what comes back,
fail in public, sharpen, repeat.

## Even when agents catch up

The autonomous alternative is arriving faster than the people waiting for it
expected.

[METR's Time Horizon 1.1](https://metr.org/blog/2026-1-29-time-horizon-1-1/),
published January 2026, measured Claude Opus 4.5 completing about five hours of
work autonomously at 50% success. The doubling time accelerated from seven
months to under three. The benchmark is starting to saturate; METR is racing to
build tasks the latest models can't beat.

Ethan Mollick, who used to call this era *"co-intelligence,"* now calls it
[something else](https://www.oneusefulthing.org/p/the-shape-of-the-thing):
*"This is an era of managing AIs, rather than working with them."* You give an
agent hours of work and review the finished product.

That's the same bottleneck shift, accelerated. *A Super Human is someone whose
bottleneck has already shifted to judgment.* The agent era doesn't change that.
It makes it the only thing that matters.

The smart voices in the field have been saying the same thing for a year.
Anthropic recommends [starting with the simplest non-agentic
solution](https://www.anthropic.com/research/building-effective-agents); agent
autonomy *"increases costs and compounds errors."* Karpathy calls for an
[autonomy slider](https://www.latent.space/p/s3), not a slider stuck at 100%.
Simon Willison spent late 2025 [designing agentic
loops](https://simonw.substack.com/p/designing-agentic-loops) carefully, not
turning them loose. The people who get the most out of these systems are the
ones who know when to grab the wheel.

That judgment gets built in the loop. The people who started running it a year
ago will be the ones managing agents well next year. Everyone else will be
reviewing output they can't grade.

## Stop waiting

Models are already good enough, and the people using them are already pulling
ahead.

You can see it already. I built [tusk](https://github.com/germanamz/tusk) over
eight weeks. The idea kept outrunning the code, until the complexity itself was
the signal. I made the call: scrap the task-management CLI I'd started with,
rewrite it as an agent-first brain. The information for that call was already in
front of me. What I needed was the judgment to trust it.

Better judgment improves how you use these models, which accelerates your
access to information, which sharpens judgment again. That loop compounds.
While others wait.