Jack Dorsey says there is no moat anymore. Yet, Palantir is up 668% in five years, Instinct is valued at $10 billion, and Emergent hit $100 million in annualized revenue in eight months. Both can’t be true unless the moat moved.

My bet: it moved to context. The software that wins from here will behave less like a tool and more like a consultant who knows your business deeply: one that tells you what to do, and then does much of it for you.

Is there still a moat?

The cost of building software is collapsing, and agents can now do real work, so everyone is asking. Jack Dorsey (co-founder of Twitter and Square) says there is none; others say distribution is the moat, and Karri Saarinen (Linear CEO) argues that individual taste will always create moats.

https://x.com/jack/status/2114121702515114234

What the winners have in common.

None of them won on the ability to build software. Technically, anyone can release an assistant; Instinct is still worth $10 billion. They won on knowing the customer’s world well enough to do the work, not just support it.

Palantir, built on embedding deeply in how each customer operates, is the clearest case:

The moats have shifted, and not everyone has caught up.

The old moats were the ability to build software, owning the customer, and feature breadth with its integrations and switching costs. Now that software can be generated, the first and the last no longer count. The new moats are:

• Depth of context. Understanding the customer's business, constraints, preferences, and history improves every decision.
• Owning the workflow. Coordinating and executing work embeds your system into everyday business operations.
• Delivering outcomes. Reliably completing valuable work makes measurable results the reason customers stay.

The difference matters. Code can be generated in an afternoon. Context can’t: it builds up from use, decisions, and outcomes, and it compounds.

You can already see the shift. Google Ads’ auto-applied recommendations adjust your campaigns based on your own account’s history, and Shopify Sidekick answers questions and takes actions using your store’s data.

Where the context lives: Plain Markdown

If context is the moat, the practical question is where it lives. The agent world has quietly settled on an almost embarrassingly simple answer: plain Markdown files the agent reads before it acts. That’s the harness: the files, triggers, and tools around a model that decide what it knows and when it moves.

Look at OpenClaw, an open-source personal agent, or gstack, Garry Tan’s open-source Claude Code setup. Each file answers one question:

• SOUL.md/IDENTITY.md: Who am I?
• AGENTS.md/CLAUDE.md: How should I work?
• USER.md/MEMORY.md: What should I know?
• SKILL.md: How do I do this task?

A chat, a schedule, or an event triggers the agent. It reads those files, then understands, acts, and checks, using tools like a browser, APIs, and code. The result is an answer or completed work.

The same harness, for an ecommerce store.

For an online store, the harness comes down to five files. What makes them work is being specific. Here’s what they might say for a skincare brand:

• Goals.md: What are you optimizing for? “Q4 net revenue of $2.4M, up 25%, with revenue at least 3× marketing spend.”
• Ground Principles.md: Which principles really matter to you as a company? “Every order keeps $18 of margin after product, shipping, and ad costs. Growth below that floor is rejected.”
• KPI Deconstructions.md: What drives each of your core metrics? “Revenue = sessions × conversion × average order value. Find which one moved before changing anything.”
• Learned Patterns.md: What have years of running the business taught you? “When the Serum is out of stock, Barrier Cream sales drop about 30%. Reorder at 21 days of stock, not 14.”
• Users.md: Who are your users, and what does each persona care about? “TikTok buyers reorder more within 60 days than Meta buyers (22% vs 15%). Judge them on 60-day value, not first-order return.”

What it looks like inside a store.

Now put a harness like this inside a real store. On day one, it simply sets honest expectations: Sam’s first sale will usually take six to seven days.

By month two, it doesn’t wait to be asked. Sam’s Goals.md says returns should be under 4% by December; they’re at 5.8%. So it brings three questions worth asking, each with a “Research this” button.

And when there’s work to do, it says so: nine listings with reduced visibility, fourteen photos to fix, three listings flagged in reviews. Next to each one is a single button: Do it for me.

What you can do today.

Start with the five files. Write down your goals, your principles, how your KPIs actually break down, the patterns you’ve already learned, and your users. Write them in plain Markdown, and keep them current.

That’s what turns software from a tool that runs a workflow into something closer to a high-powered consultant with very deep insight into your business: one that tells you what you should be doing, and then does a lot of it on your behalf.