The Agents Don't Do the Math

How a three-agent system turns SEC filings into a structured valuation config — and why the boundary between judgment and calculation is hard-coded.

When I published The First Pillar Stands last week, I mentioned that AI agents were the next step. Pillar 2 is now taking shape and the architecture is cleaner than I expected

The core principle is simple: agents author configuration. The model does the math. Never the other way around. Here is what that means in practice. When the system onboards a new company, three agents run in sequence.

1. An Analyst reads the 10-K looking at business description, MD&A outlook, risk factors, segment notes and proposes a structured configuration: drift priors per segment, FX exposures, peer multiples, capital structure thresholds. Every numerical assumption must be backed by a filing citation. No citation, no override.

2. A Reviewer then challenges each proposed field independently, cross-checking segment growth priors against historical CAGRs, verifying FX exposures are cost-ratio effects rather than translation noise, and flagging anything that falls outside acceptable bounds.

3. A Manager reconciles the two, picks the better value for every contested field with documented reasoning, and writes the final configuration file.

4. The human reviews and promotes. Agents never go to production directly.

What makes this architecture interesting is not just what the agents do, it is what they are explicitly prohibited from touching. Variables like calibrated volatility parameters, working capital days, cost of debt, beta are all computed from data, never authored by an agent. The boundary between judgment and calculation is hard-coded into the schema.

The honest limitations are in the deck too. Peer ticker hallucination is real. The reviewer can be over-cautious when management is genuinely right about forward outlook. And the self-healing XBRL loop has not yet been stress-tested on a genuinely broken filer.

Pillar 3, multi-company screening, is next, after the additional development of pillar 2 to a pilot phase.

The methodology deck is attached.