Fei Ma

Building AI that actually ships | founder | sharing real observations, not hype

[Fei Ma @xFeiMa]

Most "agent-native" rollouts don't fail on model quality. They fail on ambiguity.

"Approval is done" — by who, when? Thirty states where eight would do. A prompt that drifts while the state machine doesn't.

The agents are capable. The operating system around them isn't yet.

[Fei Ma @xFeiMa]

@miltonheyan This tracks. The people "best at agents" aren't out-engineering anyone — they've made their actual work legible to the agent: state it can read, write back to, and act on. The skill isn't prompting. It's turning your workflow into something an agent can operate.

[Fei Ma @xFeiMa]

The teams furthest ahead on agent-readable ops aren't writing smarter prompts.

They turn workflows into explicit state machines: immutable artifacts, boolean exit rules — the contract between human and agent.

What I'm seeing 👇
https://t.co/c3cELGPl1C

[Fei Ma @xFeiMa]

Most "AI company" attempts fail on the plumbing, not the model.

Bolt an agent onto a company it can't read, give it no rails, route everything (or nothing) past a human → an expensive intern.

Writing the firsthand log of what actually works → https://t.co/UxVk77gj0R

[Fei Ma @xFeiMa]

The teams where this works don't have smarter models than everyone else.

They have a better operating system around the model:
• state the agent can read + write
• guardrails it can't cross
• an approval queue for the calls that matter

Miss one and the whole thing stalls.

[Fei Ma @xFeiMa]

It shows up in three stages:

→ agent-readable: the agent sees the company's real state without asking a human
→ agent-writable: it writes back — updates the metric, files the decision
→ agent-operable: it runs standing work on a schedule, inside guardrails

[Fei Ma @xFeiMa]

Everyone's arguing about which model is smartest.

The companies pulling ahead are doing something quieter — making themselves agent-readable: an AI agent can see the real state of the business, write back to it, and run parts of it on a schedule.

Most companies can't yet. That's the gap.

[Fei Ma @xFeiMa]

@RepoGems 🔥🔥

[Fei Ma @xFeiMa]

@swyx The people who called AI trends correctly in 2022-23 were not lucky. They had a mental model of how capability curves compound. The lesson: build that model, not a prediction list.

[Fei Ma @xFeiMa]

@sama What strikes me: the people who understand software deepest are now the most effective AI collaborators. The manual mastery was not wasted — it became the foundation for knowing what to ask for.

[Fei Ma @xFeiMa]

@levelsio GPU scarcity + model capability curve = the compute supercycle is just getting started. The founders who lock in capacity now will have a structural advantage that compounds for years.

[Fei Ma @xFeiMa]

@karpathy The "phase shift in engineering" framing is exactly right. Most companies are still optimizing the old org chart with AI. The ones that win will have thrown the org chart out entirely.

[Fei Ma @xFeiMa]

Most founders treating AI as a productivity tool are thinking too small. The real play: redesign your business model around AI economics. Labor is not just cheaper — it is infinitely scalable. The constraint is your imagination, not your headcount.

[Fei Ma @xFeiMa]

Most companies don't need an AI strategy. They need to automate one annoying workflow. Ship that first. The strategy will write itself.

[Fei Ma @xFeiMa]

@naval Except the app actually ships. The podcast moment created a lot of content and very few businesses. The coding moment might be different — the feedback loop from idea to working product is weeks, not years.

[Fei Ma @xFeiMa]

@dhh The TUI aesthetic is a bet on developers who want to own their stack again. The 'vibe-coded adventures' line tells you exactly who this is aimed at — people who got tired of the SaaS subscription treadmill.

[Fei Ma @xFeiMa]

@tobi The 'reaching reflexively' is the leading indicator. The lagging one is whether it changed outcomes. One year is long enough to separate the teams who actually rewired their workflows from those who added an AI wrapper and called it transformation.

[Fei Ma @xFeiMa]

@bentossell This is the real 'AI tax' nobody talks about. You pay with compute costs, you pay with data rights, and most people don't read the terms until it's too late. The free tier is always the product.

[Fei Ma @xFeiMa]

@swyx The AI era is stress-testing all of these. Systems are being out-competed by brute-force optimization; discipline is being replaced by automated iteration. Which ones still hold when the leverage is 100x?

[Fei Ma @xFeiMa]

@paulg Same goes for anything with 'transformation' in the name. Real change doesn't need to announce itself.