Last updated: September 3, 2026
Agents61 uses large language models to power 61 AI agent personas. When you request a report, only the seats your plan unlocks run: Analyst 16 isolated prompts, Associate 29, Principal 48, Committee or Partners 61 — plus one clerk assembly pass. Masters do not read each other's drafts. Debate, when unlocked, starts after those briefs exist.
Filing facts come from SEC EDGAR (cached). Optional model-input ratios (P/E, P/B, P/S) may be cached from a licensed fundamentals API and are labeled as such — not live quotes. Crypto snapshots, when shown, come from a public market API and are cached. On-chain series (MVRV, fees, TVL) stay blank until those hooks are configured. Missing numbers stay inconclusive.
Live desks call DeepSeek V4 when a server key is configured: deepseek-v4-flash (checkpoint V4-Flash-0731) for methodology-card seats with thinking off, and deepseek-v4-pro (checkpoint V4-Pro-0813) for debate and verdict seats with thinking on. A seat that times out or fails to parse falls back to that master's methodology-card template so the desk still finishes. Stock report pages stay on the template writer so crawlers are not blocked on 16–61 model calls. Compute scales with unlocked seat count, not with pairwise debate. Models are commercially licensed. We do not use the retired aliases deepseek-chat or deepseek-reasoner.
The AI personas are simulations inspired by publicly available information about each Master's investment philosophy, published works, and documented decisions. They are not endorsed by, affiliated with, or representative of the actual individuals, their estates, or their affiliated organizations.