Agent Investment Committee
Which agents sit on the committee, how are their votes combined, and when is an agent benched, so that the committee beats the best single agent and the unscreened book after costs?
Headline result
Built a five-seat point-in-time voting committee (two news readers, an earnings-release reader, value-profitability and momentum rules) on 100 S&P 500 names, after small dated LLMs gave 0 valid votes in 200 calls. Scored once over Jan 2018 to Dec 2023, majority voting earned a Sharpe of −0.16 against 0.02 for no screen.
Figures
The two news seats are the most redundant pair at 0.45, just under the pre-registered 0.5 bloc threshold. 
Majority voting trailed throughout, ending at a Sharpe of −0.16 against 0.02 for no screen and 0.35 for the filings seat. 
The filings-lexicon seat (0.35) and majority plus benching (0.24) led; the pre-registered majority rule scored −0.16. 
Synthetic studies predicted 0.54 for majority voting; on real data it scored −0.16, because the real seats showed almost no skill.
Data used
- SEC EDGAR 8-K earnings releases (item 2.02), 4,028 releases from 99 of 100 firms, 2014–2023
- FNSPID news headlines, 3.91m after cleaning, 2013–2023
- S&P 500 prices and characteristics from the AS free route (FNSPID, Nasdaq.com, SEC XBRL) for 100 sampled members
Source paper
Caner, Capponi, Sun and Tan (2026), Designing Agentic AI-Based Screening for Portfolio Investment
Builds on Agentic AI Screening for Portfolio Investment: Replicate, Audit, Extend.
Sign in to open the full write-up, notebook, tested code and data. Papers to Projects is free with an account; no paid membership needed.