I liked this one because it gets past the usual AI chat.

Not "which chatbot should we buy?"

More like:

What would the company look like if AI was part of the operating system?

Pete Koomen helped build a lot of YC's internal AI infrastructure, and the useful thing here is how practical it all is. This is not a futurist chat about AGI. It is much more grounded than that. They are talking about databases, tools, transcripts, internal workflows, Slack channels, and the weird cultural choices you have to make if you actually want this stuff to work.

The big lesson: AI becomes far more useful when it has company context and can touch the actual machinery of the business.

Shared Organizational Brain

The strongest phrase in the whole episode is "shared organizational brain."

That is the whole thing.

The company records the important artifacts of work, gives agents useful access to them, and then lets those agents help people do better work.

Not just meeting notes. Calls, customer conversations, internal notes, product context, financial workflows, database records, skills, tools. All the messy context that normally lives in people's heads or gets buried in some SaaS product no one searches properly.

The compounding effect comes when the agent can see how the best people in the company make decisions and then reuse that pattern for everyone else.

A chatbot answers a prompt.

A company brain improves the work system.

The Boring Unlock

One of the first big unlocks was letting an agent run read-only SQL queries against YC's internal database.

That sounds boring, which is exactly why it matters.

YC has a big advantage because so much of its company context lives in one Postgres database. Companies, founders, notes, financial transactions, batches, investors. The stuff that most companies scatter across six tools and then pretend is "integrated."

Once the agent could query that database, non-technical people could ask questions that previously would have required someone else to write SQL.

The important bit is not that SQL got faster. The important bit is that people started asking more questions.

When a question takes hours, you save it for something important. When a question takes seconds, you ask all the little questions too. That changes how a company thinks.

Tools Beat Demos

The next layer is their tool registry.

This is where the agent stops being a clever analyst and starts becoming useful at work. YC began with a small number of internal tools and grew that into more than 350 tools across the company.

That matters because most company work is not just thinking. It is looking things up, updating records, scheduling, drafting, booking, checking, comparing, filing, and nudging.

The agent needs handles on the business.

For a normal company, a basic version might look like this:

• Query the internal database
• Read model files and schema
• Pull company or customer records
• Draft a customer email from meeting context
• Summarise a sales call
• Book a finance entry
• Create a reusable skill from a repeated workflow

This is the part a lot of companies will skip. They will buy the AI writing assistant and wonder why nothing really changes.

The leverage is in connecting AI to the boring internal machinery.

The Skill Loop Is The Interesting Bit

The most interesting example was the two-sentence company description.

YC partners help founders explain what their company does in two sentences. That sounds tiny, but it is one of those high-frequency judgment tasks where a lot of experience is hiding inside people's heads.

The workflow they describe is excellent:

• A partner writes an initial skill for the agent.
• Founders try writing their descriptions.
• Partners give feedback in meetings.
• The meeting transcript becomes new training context.
• The agent improves the skill from the feedback.

That is the loop.

The company takes expert judgment, captures the artifact, improves the reusable skill, and makes the next version available to everyone.

This is why the example is good. It is not "AI writes a better pitch." It is "the organization learns how it writes better pitches."

Record More Work

This is the part that will feel uncomfortable for a lot of businesses.

If you want the organisation to become AI-native, you need to record more of the work. Meeting recordings are not just about note-taking. They become raw material for improving the system.

That is a cultural change as much as a technical one.

Two years ago, recording every meeting felt strange. Now it is becoming normal. The reason is not just convenience. It is that the transcript becomes a company asset.

If your best salesperson gives unusually good feedback on a call, that can become part of a sales coaching skill. If your best product person explains a tricky roadmap tradeoff, that can become part of how new people learn. If customer complaints keep showing the same pattern, an agent can find it across calls before a dashboard catches it.

The company starts to remember.

The Culture Problem

There is a catch.

YC's version works because they have a high-trust culture. They talk about internal agent conversations being broadcast in Slack so people can learn from each other, but also so there is a bit of social control over how the system is used.

That is more important than it sounds.

Most companies will try to solve this with permissions first. YC seems to be saying that if you lock everything down too tightly, you remove much of the value. But if you open it up, you need trust, transparency, and norms.

That probably works best in a startup or a very strong team. It is much harder in a political enterprise where information is treated as territory.

Paying To Live In 2028

One of the spikier points is that a company willing to spend heavily on tokens today can temporarily live a few years in the future.

The logic is pretty clear.

AI costs will keep falling. What feels expensive now may feel normal soon. So if you can afford to spend ahead of the curve, you get the future workflow before competitors do.

That is not a reason to waste money. It is a reason to understand what you are buying.

You are not buying tokens. You are buying organizational learning.

If the company gets better at capturing context, building tools, improving skills, and letting people work through agents, that knowledge compounds. The token bill is partly an R&D bill.

Chat Might Be The Right Interface

I used to think chat was probably a temporary UI. A box you type into until someone builds the real product.

But the argument here is good: language is close to thinking. If the system can accept text, voice, images, files, and context, chat becomes less like a textbox and more like the front door to flexible software.

This connects to the later idea of just-in-time software.

Instead of building a big rigid Rails app for every workflow, you build a smaller set of tools, skills, retrieval systems, and agent instructions. Then the work can be shaped in the moment.

That is a big shift.

Traditional software says: decide the workflow, encode the workflow, make the user follow it.

Agent-native software says: give the model enough tools and context, then let the user shape the workflow through language.

Where I Would Start

If I was applying this inside a normal company, I would not start with a grand "AI transformation" program.

I would start with one repeated workflow where good judgment matters.

For example:

• Writing a two-sentence company description
• Summarising customer calls into CRM notes
• Preparing for sales meetings
• Reviewing project risks
• Turning finance questions into safe read-only analysis
• Drafting weekly leadership updates from internal context

Then I would build the smallest possible loop:

• Capture the work artifact
• Let the agent use relevant context
• Add one or two useful tools
• Watch the best person do the task
• Turn their feedback into a reusable skill
• Improve the skill every week

That is the playbook. Not one big AI tool. A small loop that gets smarter.

Why I Liked It

The reason I liked this episode is that it makes AI feel less like magic and more like company design.

Most organizations are full of trapped context. People know things. Meetings produce decisions. Customer calls reveal patterns. Finance teams answer useful questions. Product leaders have taste. Salespeople have instincts.

But most of that knowledge disappears after the moment passes.

YC's AI playbook is an argument for making that knowledge reusable.

That is the real prize. Not a chatbot. Not a prompt library. Not a novelty demo.

A company that can remember, reason, and improve how it works.