Karpathy for Finance & Ops
Andrej Karpathy (founding member of OpenAI and astute namer of things) gave a talk at Sequoia Capital called “From Vibe Coding to Agentic Engineering.” It was geared toward software developers, but his discussion of context and verification is useful for finance and ops too.
A model needs the relevant business context to do useful work. We also need a way to check what it produces. In finance, that means thinking about:
- The information we supply.
- The systems the agent can use.
- The checks a deliverable has to pass before someone relies on it.
Models are unevenly capable
Karpathy describes frontier models as having jagged intelligence: strong capabilities in some areas, with surprising gaps nearby. You find out where those gaps are by using the model. Success on one task gives you only so much confidence about the next.
For finance, being good at analyzing public filings doesn't establish that a model understands your multi-entity consolidation, revenue recognition policy, or chart of accounts. Those depend on information specific to your business. And even when a model produces a good answer, you still need to check the next one.
Give the agent the context and systems it needs
In Karpathy's Software 3.0 framing, the model is the interpreter and the context window is where the program lives. Think of Excel's formula engine running the formulas you write. With a language model, the instructions and information you supply shape what it does.
For a finance team, that context includes:
- Company objectives.
- Historical data.
- KPI definitions.
- Accounting policies and workpapers.
- The financial model.
The current task may also need:
- A contract.
- A forecast scenario.
- The latest board materials.

The diagram separates this into business context, a coding harness, the model and connected applications, and verification.
The coding harness is the environment where you give the agent a task and it works through it. A board memo, forecast update, or variance analysis needs both a clear request and access to the relevant context.
The agent also needs connections to the systems where the work happens: Google Sheets, the general ledger (e.g., QuickBooks, Campfire, NetSuite, Rillet), Slack, etc. Tools exposed through APIs allow it to retrieve information and take actions. Which connections it needs depends on the task.
For example, a variance analysis could involve:
- Pulling the financial statements from the general ledger into Google Sheets.
- Reading an existing forecast that lives in a separate forecast model.
- Considering business context from transcripts of recent team meetings discussing the forecast.
- Calculating formula-driven variances.
- Drafting v1 commentary, using the transcripts and general ledger transactions as context.
Check the work before relying on it
The checks should reflect the deliverable. A financial model needs reconciliations and source traceability. A board memo also needs review of the explanations and conclusions drawn from the numbers.
Write down the standards a senior teammate would apply: accuracy, completeness, consistency, explanations of variances, and clarity. Give the agent examples of acceptable work where those are useful.
Instructions alone don't ensure those standards are met. Where a check can be expressed in code, such as whether totals reconcile, it can run against the output. Agent review and human review can address the parts that require judgment. A reconciliation passing doesn't establish that the explanation of a variance is sound.
The person responsible for the work still needs to understand it. A CFO has to stand behind the board materials, including the assumptions and conclusions, even when an agent helped prepare them.
Keep the verified work for next time
Once a close package is reconciled or a forecast scenario is adopted, save it with the context the agent can use on subsequent tasks. The same applies to decisions and explanations that would otherwise be lost between runs.
A completed close package gives the next cycle:
- A target format.
- A set of controls and quality gates to clear.
This can reduce the work of reconstructing context and approval thresholds each time.
Technology stack links
Tools used: links to the official repos and websites.
Subscribe for thoughtful, cutting-edge rants from Gary
and the opportunity to beta-test new finance superpowers and parenting superpowers