one prompt · research to forecast / DeerFlow × OASIS × Graphiti

Ask one question.
Auto-research, build a world, simulate the future.

A deep-research agent (DeerFlow) searches the web and builds an evidence-grounded dossier; the system constructs a local temporal knowledge graph (Graphiti), generates digital personas for the real-world actors it found, runs a multi-agent population simulation (OASIS, dual-platform), and a report agent synthesizes the forecast. The pages below show every stage of real runs, unedited.

View the demo forecasts ↓ Run it yourself

Live demos — real end-to-end runs

Each card is one full pipeline run. Click through to walk the whole workflow: deep-research log → research dossier → ontology → knowledge graph → simulated forum → final forecast.

● completed · full pipeline中文

Storage semiconductors — the 2027–2028 outlook

“针对 2027—2028 年存储半导体市场整体行业格局开展全面前景预判,完整包含四大核心模块:厂商竞争格局、新一代存储技术、供需与价格走势、行业历史周期规律……”

40-round dual-platform simulation · 80 personas · 213-node knowledge graph · competitive landscape, next-gen tech (HBM4E / DDR6 / CXL 4.0), supply-demand & pricing, 40-year cycle history

Walk through the run →
● completed · full pipeline中文

Global cloud computing — the 2030 endgame

“你作为全球云计算产业资深分析师,综合算力储备、客户规模、区域份额、毛利率、技术壁垒、地缘合规、AI 算力增量需求等指标,完整推演至 2030 年全球云计算竞争格局,明确赢家与失利厂商……”

Dual-platform population simulation · 80 personas · 130-node knowledge graph · the Big Three vs. Oracle OCI, sovereign cloud & AI-capex dynamics — winners & losers to 2030

Walk through the run →
● completed · full pipeline中文

Who dominates US AI by 2030?

“预测到2030年,哪家美国人工智能实验室或企业将成为美国AI领域竞争的主导者……”

40-round dual-platform simulation · 42 personas · 6-section forecast

Walk through the run →
● completed · full pipeline中文

Global EV industry through 2035

“预判2035年前全球电动汽车产业的发展趋势:技术路线演进、产业链竞争格局、各国政策走向……”

40-round dual-platform simulation · 16 personas · technology / supply-chain / policy scenarios

Walk through the run →
● completed · full pipelineEN/中文

How and when does the Russia–Ukraine war end?

“How and when will the Russia–Ukraine war end? Research the current state of the conflict…”

36 personas · multi-scenario endgame analysis grounded in a 40K-char research dossier

Walk through the run →
● completed · full pipeline中文

Global semiconductors through 2030

“分析 2030 年前全球半导体行业的发展走向。围绕设计、制造、封装测试、组装等全产业链环节展开研究……”

40-round dual-platform simulation · 115 personas · full value chain: memory / HBM / logic / foundry across 17 named companies

Walk through the run →
● completed · full pipeline中文

Global memory-chip market through 2030

“预测到 2030 年全球存储半导体(DRAM、NAND 闪存、HBM 高带宽内存)市场的竞争格局:主要厂商的份额演变、技术路线、AI 驱动的 HBM 需求、地缘政治与出口管制的影响……”

4-round dual-platform simulation · 80 personas · DRAM / NAND / HBM competitive landscape — Samsung, SK Hynix, Micron & YMTC, AI-driven HBM demand and export controls

Walk through the run →
● completed · full pipeline中文

China's energy storage & battery market in 2035

“请基于以下要点,对 2035 年中国储能与电池市场做出深度、结构化预测与竞争格局研判……”

40-round dual-platform simulation · 94 personas · grid / C&I / home storage segments, full supply chain & winners-vs-losers analysis

Walk through the run →
● completed · full pipeline中文

How does the 2026 US–Iran war end?

“请基于当前地缘政治现状、美伊双方核心战略诉求、军事实力对比、地区盟友体系、大国博弈背景……开展系统性、多层级、场景化深度预测分析……”

40-round dual-platform simulation · 135 personas · endgame scenarios, peace-deal terms & post-war Middle East order

Walk through the run →

47-second demo

One prompt → research console → knowledge graph → live simulation feed → forecast.

Screenshots

Knowledge graph
Temporal knowledge graph built from the research dossier
Research dossier
Research dossier — every claim grounded in cited sources
Agent personas
Digital personas generated for each real-world actor
Simulation console
Live simulation console streaming agent actions
Simulated feed
The simulated Twitter/Reddit feed mid-run
Simulated posts
Emergent discussion threads between personas

How it works

  1. research — DeerFlow deep-research agent searches the web, extracts key actors (role / stance / influence) and writes a cited dossier
  2. ontology — an LLM derives entity & relation types from the dossier and your question
  3. graph — the dossier is ingested into a local temporal knowledge graph (Graphiti, GraphRAG)
  4. prepare — researched actors become digital personas with evidence-based stances
  5. run — hundreds of LLM personas interact on a simulated Twitter + Reddit (OASIS)
  6. report — a tool-augmented ReAct agent queries graph + simulation and writes the forecast
# run it yourself — no graph DB to host (embedded Graphiti/FalkorDB);
# all you need is a Claude/Codex CLI login or any LLM API key
git clone https://github.com/linroger/DeepAgentForecast.git
cd DeepAgentForecast && ./setup.sh   # interactive: picks your LLM provider
npm run dev   # → http://localhost:3000/research