Skip to content

Repository files navigation

Agent Valley 🏔️

An open-source AI startup ecosystem simulator — a multi-agent LLM-driven what-if scenario sandbox for investors, VCs, and strategists.

30 AI companies — foundation labs, startups, VCs, big tech, wildcards — autonomously compete across 8 market segments (enterprise AI, consumer AI, healthcare AI, developer tools, cloud infrastructure, AGI research, robotics AI, fintech AI). A deterministic economic engine ensures mathematical rigor; LLMs drive strategic decisions. Extensible to any industry via custom scenario files.

License: MIT Python 3.10+


Why Agent Valley?

Venture capital and corporate strategy decisions are made with backward-looking data (CrunchBase, PitchBook) and gut feeling. There is no interactive sandbox where you can ask:

"What happens if the EU passes a new AI regulation?" "If Google acquires Anthropic, who wins the enterprise AI market?" "Which AI infrastructure startups survive the next 2 years?"

Agent Valley answers these questions with simulation, not speculation — an LLM-powered AI industry competitive landscape simulator that turns "what-if" into "here's what the data says."


Quick Start

git clone https://github.com/emergencescience/agent-valley.git
cd agent-valley
python -m venv .venv && source .venv/bin/activate
pip install -e .

# Deterministic-only (no API key needed, instantaneous)
python scripts/run_sim.py --quarters 4 --no-llm

# Full simulation with LLM strategic decisions
export LLM_API_KEY="sk-..."
python scripts/run_sim.py --quarters 8 --seed 42

# Use OpenAI instead
export LLM_API_KEY="sk-..." LLM_BASE_URL="https://api.openai.com/v1" LLM_MODEL="gpt-4o"
python scripts/run_sim.py --quarters 8

How It Works

7-step quarterly pipeline:

① Market Pulse    → Fed rate, segment TAM growth, VC sentiment
② Agent Decisions → LLM: each company chooses strategy (pricing, R&D, hiring, M&A targets)
③ Economy         → DETERMINISTIC: revenue, burn, cash, market share, valuation
④ Deal Matching   → LLM: partnerships, acquisitions, funding rounds (validated by engine)
⑤ Guardrail       → Hard/soft constraints — violations trigger reversion
⑥ Black Swan      → Probabilistic: regulation, market crash, tech breakthrough
⑦ Narrative       → LLM: fact-based quarterly report

Core Design Principle

Agents do NOT communicate directly via LLM chat. They interact through shared market signals managed by the deterministic economy engine. This avoids O(n²) complexity, LLM hallucination in negotiations, and makes outputs empirically validatable.

Layer Responsibility Implementation
Deterministic Economy Revenue, market share, valuation, funding Pure math (softmax, P/S ratios, TAM × share)
LLM Strategy Pricing, R&D allocation, partnerships, M&A targets LLM with structured JSON output
Macro Environment Fed rate, regulation, geopolitics, breakthroughs Parameter sweep + black swan injection

Example Output

🏆 2026 Q4 — Market Cap Leaderboard
#  Company            Valuation     Cash      Rev/q    Status
1  Apple              $1,463.6B    $30.1B    $1.8B    public
2  Microsoft          $1,291.4B    $23.3B    $1.9B    public
5  DeepMind           $165.4B      $19.2B    $3.0B    private_subsidiary
6  Anthropic          $75.9B       $14.3B    $1.5B    private
7  OpenAI             $50.2B       $28.5B    $2.7B    private

Active: 28 | Acquired: 2 | Bankrupt: 0 | Deals: 41

FAQ

What is Agent Valley?

Agent Valley is an open-source, MIT-licensed multi-agent LLM-driven simulation of the AI startup ecosystem. 30 AI companies compete across 8 market segments in quarterly cycles. A deterministic economic engine computes revenue, market share, and valuation while LLMs drive strategic decisions (pricing, R&D, partnerships, M&A). Outputs include time-series leaderboards, automated narrative reports, and investment theme identification.

Who is this for?

Agent Valley is built for investors, VCs, corporate strategists, and policy advisors who need to answer "what-if" questions about the AI industry: Which segment will dominate in 2 years? What happens if regulation changes? Who should I acquire — and when? It is also a research tool for academics studying multi-agent economic simulation and generative agents.

How is this different from PitchBook or CB Insights?

PitchBook and CB Insights tell you what happened. Agent Valley simulates what could happen. It's an interactive sandbox where you can change parameters (Fed rate, regulation strictness, tech breakthroughs) and see how the competitive landscape evolves. It's complementary to data platforms, not a replacement.

Can I simulate industries other than AI startups?

Yes. Agent Valley is extensible via scenario files (scenarios/{name}/agents.json + calibration.yaml). Define any set of companies, market segments, economic parameters, and archetypes. The simulation engine is domain-agnostic — the default scenario happens to be the AI startup ecosystem because that's where the most urgent investment questions are.

Does it require an LLM API key?

LLM mode requires an API key. Set LLM_API_KEY (or OPENAI_API_KEY, or the legacy DEEPSEEK_API_KEY). Works with any OpenAI-compatible endpoint — DeepSeek (default), OpenAI, Anthropic via proxy, Ollama, vLLM, Groq, Together, etc. Configure via LLM_BASE_URL and LLM_MODEL env vars. Default model is deepseek-v4-pro for maximum accuracy — the simulation is designed for insight quality over speed.

Deterministic mode (--no-llm) runs without any API key.

How accurate are the simulations?

The deterministic economy engine uses empirically-grounded formulas (softmax market share, P/S ratio valuation, GPU supply constraints). LLM-driven strategic decisions add qualitative reasoning about competitive dynamics. We are building an empirical validation framework (backtesting against CrunchBase survival rates) — see docs/architecture.md. Currently, simulations should be treated as structured thought experiments, not financial advice.

How do I contribute?

See CONTRIBUTING.md. The most impactful contributions are: new industry scenarios, improved economic calibration with real market data, and additional black swan events with documented probabilities.


Documentation

Document Description
Architecture System design, data model, design decisions, module map
Product Spec User stories, feature specs, acceptance criteria
Strategy Market research, user personas, business model, GTM
Development Installation, CLI, configuration, monitoring, benchmarks
GEO Guide Generative Engine Optimization playbook for maintainers
Contributing Setup, code standards, PR process, how to add scenarios
Changelog Version history and release notes
Security Vulnerability reporting

Research


Scenarios

Scenario Agents Segments Description
silicon-valley 30 8 Default: AI startup ecosystem (2025 baseline)

Create custom scenarios in scenarios/{name}/agents.json + calibration.yaml. Extensible to biotech, climate tech, crypto, or any competitive industry.


Ecosystem

Agent Valley is part of Emergence Science — a platform for agent-to-agent economy.

Related projects:


License

MIT © Emergence Science


"The simulation is a means; the investment thesis is the product."

About

Open-source MIT-licensed multi-agent LLM-driven simulation of the AI startup ecosystem — a what-if scenario tool for investors, VCs, and strategists

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages