Equities, options, bonds, crypto, prediction markets. Voice-driven backtests, fact-checked theses, on-demand quant pipelines — orchestrated by agents on production-grade infrastructure.
Source — a research-engineer workflow describing what to surface daily: scheduled prices / market caps / volumes for major assets, breaking news, and AI supply-chain signals (chips, memory, GPU shipments). The agent runs the loop on schedule.
Agent output — ranked signal packets pushed to Telegram and the dashboard. Each links a price move to a news catalyst. Includes the AIMEM memory-price index as a real-time AI-infra leading indicator.
Traders and researchers describe hypotheses to the agents — by voice or text. The agents formalize each into a backtested strategy across equities, options, bonds, crypto, and prediction markets. Three real cases from anonymized design partners below — each verified and validated by real market data and live computations.
Why agentic matters — a traditional backtest workflow is single-threaded: a team of researchers + engineers, weeks of meetings. The agentic factory runs multi-threaded backtests in parallel, anytime. Each strategy is an extension of the trader's life — every fleeting intuition gets validated, not just the few you have time for.
Source — coverage of the global AI infrastructure stack: 1,590 earnings-call transcripts + 1,445 disclosure events + 26 Annual Report MD&A digests, across 38 stocks in 6 countries (US, China HK/ADR, China A-share, Korea, Taiwan, Japan, Netherlands) and 10 industry layers. The agent reads everything, runs Deep Think reasoning, writes a fact-checked long-form thesis with verbatim citations.
Agent output — a position-sizable investment thesis. Every numeric claim traces to a specific management quote with date attribution. Includes cross-Pacific verbatim comparison, time-series evolution across 6 quarters, and falsifiable trade ideas — each with the specific event that would force thesis reversal.
Hallucinated a Cambricon 690 deployment that does not exist; conflated iFlytek's Ascend roadmap with shipped product. Headline figures borrowed verbatim from a single Tom's Hardware aggregate.
Bottom-up: per-company live fetch, every figure recomputed. Held back at the paywall — Tencent GPU allocation verbatim, niche litigation events, cross-corpus mention rates all out of reach.
Owns the corpus the others can't reach: paywalled global earnings transcripts, MD&A filings, A-share / HK disclosures. Cross-Pacific verbatim quote comparison, 6-quarter time series, falsifiable trade triggers.
Net read — general-purpose deep-research tools are now comparable to each other. The corpus is the next tier. The moat isn't the model. It's the data the agent can actually read.
Pick the right method for the question. Agents handle the full ladder — from classical statistics through deep learning and reinforcement learning to large language models — and orchestrate the systems to run them.
Orchestration — all driven by an LLM agent layer that allocates CPU and GPU, manages memory and disk storage, and runs the full pipeline end-to-end — from raw ingestion to a backtested report.
What changes when the factory is agentic.
Four shifts that didn't exist a year ago — each expands what a single trader can do.
A strategy idea used to mean meetings and weeks of code. One voice memo now becomes a backtest in minutes. Inspiration that would have died in a notebook gets validated.
Factor decomposition, walk-forward validation, ML pipelines — agents do them on demand, by description. The specialist apprenticeship is no longer the bottleneck.
LLMs read and reason over news, filings, transcripts, research papers — multi-threaded inference at a scale no human team could match. Signals embedded in language, not just numbers, become tradeable.
One human watches one screen. Sub-agents watch dozens of order books, news feeds, and curves in parallel. Signals get aggregated; the trader reasons over what was found.
Agentic Sciences is a time-production company. The agent is an extension of life — it reads more, watches more, attends more, on your behalf. You get the hours back.
The same agent stack powers our research engine. What it has already shipped:
Taking investor contacts now
Qihong Ruan · Cornell Econ PhD '25
Ollama × Google DeepMind · community event with Gemma 4 release. LinkedIn post →
Google DeepMind × Stanford 2026 Hackathon · built Market Archive — the trader's time-travel (Lehman, COVID, 2024 yen unwind). Live app → · LinkedIn post →