The Great Convergence | Nicholas Charriere The Great Convergence

Over the last year, a strange thing has happened in tech: very different companies have started moving towards the same product shape, and it feels like everyone is building the same thing.

Linear announced last week that they’re building coding agents. OpenAI is deprecating Sora and focusing entirely on Codex. Anthropic is obviously all-in on claude code and cowork. Notion is building agents for work. So are Google, Microsoft, Meta (Manus), Lovable, Retool and many others.

These companies have different histories, customers and product categories, but they’re starting to converge on the same idea: software that can take a goal, use tools, and do work on your behalf.

This convergence is not hard to explain: the market is enormous . This is so much more than a new feature. The prize is enterprise knowledge work.

What changed

The important shift is not just that models got better (although that’s a major part of it working), it’s the invention of the general harness .

Claude Code was a massive breakthrough. Although initially invented for coding use cases, it turns out that a smart looping agent generalizes incredibly well towards any computer based task if you give it the right tools. (claude code → claude cowork)

So this new technique emerges and turns out to be a general problem solving machine . It also scales on a very unique dimension: it can keep running for a long time (autonomy level is theoretically a configuration of the system).

This is the holy grail of software: a conceptually simple system that can solve many problems.

It takes the shape of a model harness + a goal + a set of tools. It runs in a loop calling tools until it stops and produces a result.

I’m simplifying a little and there is a lot of execution craft in doing these things well, but if you build this and then throw tokens at it there seems to be very little theoretical ceiling to what it can achieve. More tools, more tokens, more greedy approaches all seem to scale, and models will only ever get better from this point out.

In the last couple of years, it has become evident that LLMs are incredible at writing code. They’re so good that coding is the current best assumed path to AGI. And these harnesses as well as their tools are all just code.

Because the harness + model = code + intelligence, they have the ability to reflect and improve themselves over time. Andrej Karpathy went viral with this idea in his recent prototype: AutoResearch.

The opportunity

The prize is not one more AI feature; it’s automating enterprise knowledge work. If you can sell this to other companies, you’re selling labor itself. There is probably no ceiling to the demand.

This is such a large opportunity that it makes “personal AI use cases”, the B2C motions like Sora or chatGPT, seem trivial. In this light, it makes perfect sense for openAI to refocus away from B2C to focus on this. Everyone should, it’s the economically rational thing to do.

If you can sell a tool like Claude Cowork or Notion Agents or “Manus for enterprise”, you essentially are selling generic business outcomes as a service.

Continuous Learning

When I was working on self-driving cars at Cruise, the vision was the concept of a “Continuously Learning Machine” (CLM). The goal was to have the cars drive around, gather experience and then improve, and it guided all of engineering’s efforts.

This was of course never quite achieved, there were always humans in the loop. It was nonetheless a very useful north star: over time humans steered less, and they did so only in the highest leverage areas (tough labeling, model tuning, deploy decision).

We massively compressed the feedback cycle, and in the 4 years I was there we went from deploying new models on the car ~quarterly to ~weekly. That pace of iteration was incredible and led Cruise to deploying autonomous vehicles first in SF, ahead of Waymo.

This idea of continuous learning is about to spread everywhere in these agentic products. Building software factories that are as autonomous as possible is happening everywhere. Just like the CLM, humans will always be somewhat involved.12

I think that a lot of the competition will be across this dimension. Different products will be on different places of the autonomy slider.

At Cruise, the feedback loop was drive → collect data → retrain the model → deploy. For LLM agents, the loop is run → monitor → improve the harness code and context engineering → run again. The difference is that the agent itself can close this loop. It can reflect on its own performance and use its coding ability to implement better approaches.

Stanford Researcher @yoonholeee recently announced exactly this idea, that he coins as “Meta-Harness”:

The app layer convergence

The pattern is everywhere:

an agent to a calendar app

an agent to a travel booking app

an agent to a house listing website

an agent to a maps app

an agent to an email...