Knowledge packs and the Software 3.0 shift

Andrej Karpathy — one of the founders of OpenAI and the former head of AI at Tesla — recently described a shift he calls Software 3.0. The idea is simple: in 1.0, we wrote code. In 2.0, we trained models. In 3.0, we prompt them.

He uses a computer analogy to explain it. An AI model has two parts: the weights — everything the model learned during training, baked in permanently, like a brain's long-term memory — and the context window — the short-term working memory where it holds whatever you give it right now. Karpathy's point is that the weights are the processor (they do the reasoning) and the context window is the RAM (it holds the information you're working with). Programming in this new paradigm isn't writing code — it's deciding what goes into that working memory.

When I heard that, I realized it's a precise description of what Workflow People already does — and knowledge packs are the mechanism that makes it work.

How knowledge packs map to Software 3.0

Every time you run a command in Workflow People, the app loads your knowledge packs into the AI's context window before it does anything. Your chart of accounts, your business rules, your API schemas — they become the "program" the AI executes against.

The AI does the thinking — categorizing transactions, building workflows, generating reports. But the understanding comes from you, encoded in knowledge packs. That's the division of labor Karpathy describes: we can outsource thinking to AI, but we can't outsource understanding.

Vibe coding vs. agentic engineering

Karpathy draws a line between two modes of working with AI. Vibe coding raises the floor — anyone can type "build me an expense tracker" and get something functional. Agentic engineering raises the ceiling — professionals orchestrate AI with precision to produce reliable, domain-correct results.

In Workflow People, the same spectrum exists:

• Day one: You type "categorize my bank transactions" and the AI does its best with generic knowledge. That's vibe coding.
• Day thirty: You've built knowledge packs for your chart of accounts, your vendor rules, your fiscal year conventions. Now the AI builds workflows that match your exact business. That's agentic engineering.

The skill you're developing isn't coding. It's articulating your domain knowledge in a way the AI can use. That's the new literacy.

Why accounting is the right domain

Karpathy points out that current AI models are "jagged" — they excel in domains that are easily verifiable, because those domains let the models be optimized through reinforcement learning. Code and math are the obvious examples.

Accounting sits squarely in that sweet spot. Numbers must balance. Categorizations follow known rules. Reports have expected formats. Double-entry bookkeeping is inherently self-verifying. Knowledge packs make it even stronger by constraining the AI's output space — "here are the valid accounts, here are the rules" — turning a fuzzy request into a verifiable one.

Knowledge packs as skills — a concrete example

Say you need currency conversion in your workflows. In a traditional app, you'd look for a "currency connector" in some marketplace, configure it, and hope it does what you need. In Workflow People, you teach the AI how to do it:

1. Settings: Add your API key for a currency service (e.g., ExchangeRate API). This is a credential — private, never exported.
2. Knowledge Packs: Create a pack describing the API — the endpoint URL, how authentication works, what the response looks like. Something like: "ExchangeRate API at v6.exchangerate-api.com. Pass the API key in the URL path. Response includes a conversion_rates object with currency codes as keys."
3. Any workflow: Just say "convert my expenses from USD to EUR using current rates." The AI reads your knowledge pack, knows how to call the API, and builds the right tool automatically.

There's no connector to install. No integration page in the workspace. The knowledge pack is the skill. It lives on the Knowledge Packs page alongside your chart of accounts and business rules — because teaching the AI about a currency API is the same kind of act as teaching it your fiscal year conventions.

This also means skills are shareable. When you save a .wfp file, the currency knowledge pack is embedded in it. A colleague who opens that file gets the workflow and the knowledge of how to call the API. They just add their own API key in Settings. The skill transfers. The credential doesn't.

The .wfp file is a portable skill

That currency example is one case, but the pattern scales to everything. A .wfp file contains your workflows, your tools, your data, and your knowledge packs. It's not just a save file — it's a complete, portable encoding of how your business works. Hand it to a colleague and they can run your month-end close with fresh data. That's Software 3.0 in a file.

Karpathy and others talk about "rebuilding the internet for agents" — moving documentation from human-readable to agent-executable. Knowledge packs are exactly that. They're agent-readable documentation of your business processes. Not a human SOP buried in a shared drive — a living document the AI reads every time it builds or runs something.

What this means for what I'm building

The value is in the accumulated domain knowledge that makes the AI accurate for your business. The workflow stays the same: describe your work, let the AI build it, run it on real data, refine it. But every knowledge pack you add — whether it's your chart of accounts, your vendor rules, or a currency API — makes every future operation more precise. Your understanding compounds. The AI's thinking gets better. That's the loop.