A language of its own, not a bigger model: one startup's recipe for keeping AI agents from getting lost in Excel

By Zendoric · October 6, 2026.

Luke Harris, co-founder of Modeloptic —a financial modeling software startup (the discipline of building projections of a company's revenue, costs and cash, usually in Excel)— has published a technical argument on the company's blog with a very concrete example. As he describes it, an AI agent (a system that does not just answer, but carries out actions on its own) that tries to write a row of a financial model by operating directly on an Excel sheet has to string together nine separate decisions: locating the right sheet and row, mapping which column corresponds to which quarter, identifying the annual total columns, recognizing the financial pattern it needs to apply, deciding which cells take fixed values and which take formulas, and building each cell reference by hand. With the domain-specific language (DSL, a pared-down programming language designed for one specific problem rather than a general-purpose one) that Modeloptic has built for its own product, the same task is solved in three steps: recognizing the pattern, remembering that a tool exists to execute it, and calling it with the right arguments.

The underlying diagnosis the article puts forward has more reach than the financial case. Harris describes two structural limitations of today's language models: they do not retain memory between conversations unless specifically trained to, and their performance degrades the more information —including what the model itself generates while "thinking"— they have to hold in the context window (the span of text the model can process at once). Hence the practice, increasingly cited in the sector, of "context engineering": the point is not to make the model bigger, but to cut incidental complexity —the friction that does not belong to the real problem— as far as possible before it reaches the model. The recipe the article proposes is to delegate to deterministic code everything that does not require judgment, and reserve for the model only the decisions that do: not exposing it to the email-sending protocol if what it needs is to send an invoice, not letting it touch a SQL database if what it needs is to update a status.

This is worth reading with one caveat: it is a product piece. Modeloptic sells exactly the platform it describes, and phrases like "AI financial modeling finally works" are a commercial claim by the company itself, not a finding verified by third parties —the original post had barely three points and zero comments on Hacker News, a sign that it is a niche technical note, not an announcement with reach. That said, the pattern it describes is neither exclusive to finance nor of their invention: it fits what Anthropic and other labs have spent months insisting on in their own agent design guides —narrow, well-documented tools beat agents with generic, unrestricted access— and with the race to control the agents' "plumbing": the competitive advantage is shifting from the smartest model to whoever best designs the layer that mediates between the model and real work.

Our take is that this piece, modest in scope, captures well where agentic AI stands. The bottleneck is now almost never the model's raw capability —frontier models are more than enough for tasks of this kind— but the reliability with which that capability translates into a correct, repeatable and auditable action. The article itself points to a secondary advantage that strikes us as more important than the saved steps: if an agent can only act through a closed catalog of operations, every action is recorded as a traceable change, not as an opaque edit to data. That is governance applied to day-to-day work, not to model geopolitics, and it is exactly the kind of boring engineering —well-thought-out abstractions, audit logs, clear limits on what the model can touch— that turns an impressive demo into a tool a company can trust with its accounting. In the short term, that software craftsmanship is what decides which administrative tasks are automated now and which still need a human to review the spreadsheet; in the long term, it is precisely this kind of reliability engineering —unglamorous, not remotely viral— that will have to mature across thousands of different domains for the promise of abundance underpinning AI to stop being a projection and become infrastructure you can genuinely count on.