How Edgar Agent works
Edgar Agent is a deterministic pipeline wrapped around non-deterministic models. A filing enters at one end, gets decomposed through four analytical lenses, and exits as a signed invest/pass signal with a confidence score and a critic review.
SEC EDGAR supplies 10-K filings. We chunk and embed them, then run four parallel lenses, synthesize a proposal, and have a separate critic agent review it before a final signal is emitted.
Extracts material risks from Item 1A. Flags concentration, leverage, legal, regulatory, and macro exposures.
Parses MD&A and business description for growth narrative, market share moves, and forward guidance.
Compares current filing against prior years. Surfaces changes in tone, new risks, and trend reversals.
Pulls quantitative signals from statements: margins, cash conversion, leverage, capital allocation.
- 1Extraction. Filings are pulled from EDGAR and sectioned by item.
- 2Chunk + embed. Text is split into ~1k-token chunks and embedded into pgvector.
- 3Lens analysis. Each lens queries the vector store and asks the LLM for structured, typed signals (numeric or categorical) with a confidence score.
- 4Synthesis. The synthesizer agent weighs lens outputs and produces a research note.
- 5Critique. A separate critic agent scores the note for rigor and missing evidence.
- 6Signal. If the critic score clears the threshold, the note is finalized with an invest or pass signal and an aggregate confidence.