Garry Tan says systems of record must become AI harnesses or face replacement by agents. This post traces the constraint from the sole practitioner's context problem, to the fifty-person practice's consistency problem, to the audit trail a regulated firm cannot do without.
On Monday, Garry Tan (of Y Combinator) gave systems of record two options: become an AI harness, or be replaced:
Prediction: systems of record will need to become AI harnesses or face replacement by agents
— Garry Tan (@garrytan) August 24, 2026
He and his company have been banging this particular drum for a number of months:
It's clearly an idea that is generating a lot of interest, so it's worth unpacking.
Large Language Models (LLMs) are generalists. They have read the public internet. They have not read your project records, your staff capabilities, or know which client always pays late. So, while an LLM can produce incredible documents and slide decks based on general public knowledge, you need to give it the "context" of your own business for it to produce the really useful stuff.
An AI "harness" is an interface that lets a human interact with an LLM. At the moment, the most high-profile examples of these are chat interfaces. But they are becoming more sophisticated - Claude Cowork and ChatGPT Work still expect you to chat with the LLM, but they can read files from your computer, search the web, and take control of other applications on your computer. For complex tasks, they can strategise - build a plan, send out sub-agents to execute bits of it, and pull together the results into a coherent output.
For a solo/small business, these commercial harnesses are revolutionary. A bunch of text files can contain all the context the AI needs to produce meaningful work that is targeted to you. I keep a Second Brain-style repository of markdown files in Obsidian, and Claude Cowork uses that as context. It means that when I ask Claude to analyse an academic paper, summarise a YouTube video or help me plan a blog post it knows exactly what my business does, what I have been reading recently, and who I am targeting.
For solo/small businesses, the main thing stopping them getting the most out of AI is resources - they lack the time or people to set up a Second Brain, GBrain or something similar. But the tools and advice are freely available, when they get round to it.
⇝
“The constraint does not disappear as a firm grows. It migrates. For a sole practitioner it is time. For a fifty-person practice it is consistency. For a firm with an audit committee it is accountability.”
The constraint does not disappear as a firm grows. It migrates. Imagine ten engineers all producing design reviews. Left alone, each develops their own prompts, standards and shortcuts. Six months later nobody can explain why similar projects receive different recommendations. Consistency is now the main problem. This is where shared standards are needed, something that cannot really be done with an ad-hoc collection of markdown files on each user's personal drive.
And as this problem is solved and businesses continue increasing in size, the constraints continue migrating. Unverified AI output, ungoverned data collection, and unaccountable decisions turn into audit exposure, compliance risk, and reputational damage. At this stage, your "Company Brain" needs controls, an audit trail, and integration with risk frameworks.
That's not to say that these constraints are being ignored. Some of them are raised in the comments below Garry's post:
Even so, there are all sorts of other benefits in standardising your staff's access to AI agents via a harness:
Despite the hurdles, it's clear that company-specific harnesses, Claude Cowork-style interfaces with instant access to all of the company's data and preferred outputs, are on the way. As Harrison Chase of LangChain recently put, a harness has one job - putting the right context into the window at the right moment. General-purpose harnesses are tuned for general-purpose work. By all means, buy the harness that drafts your emails. But build the one that knows which client always pays late.
CLAUDE.md or AGENTS.md) with three tiers (Canonical, Domain, Archival) which tells the AI which documents are source of truth, which are domain-specific references, and which to ignore unless explicitly asked.
❝
“Fighting with a large army under your command is nowise different from fighting with a small one: it is merely a question of instituting signs and signals.”
If you would like to find out more about working effectively with AI, please do get in touch.