cyberagent 0.1.4

Physical-bottleneck, reverse-consensus investment analysis for every market — an LLM agent chain that traces any stock (A-share / HK / US) to its supply-chain constraint and refuses to chase narrative-driven tops. Bring your own LLM key.

🧠 cyberagent

Physical-bottleneck, reverse-consensus investment analysis — for every market

A chain of LLM agents that traces any asset down to the physical constraint that caps its industry, checks whether the market has already priced it, and refuses to chase a narrative-driven top. A-share / HK / US stocks. Bring your own LLM key.

🌐 Language / 语言

English  |  简体中文

---

What makes it different

Most open-source "AI analyst" frameworks ask "is this a good company?" and return a textbook SWOT. cyberagent asks a sharper, falsifiable, reverse-consensus question, in a fixed order:

Physical bottleneck → uniqueness → commercialization → financial elasticity → consensus correction

Where is the physical constraint in this asset's supply chain? Is it unique? Can it be monetized? Does it have non-linear financial elasticity? And has the market already priced it in?

It is built on one idea from Leopold Aschenbrenner's Situational Awareness: AI scaling is a massive industrial process, bottlenecked by physical inputs — power, transformers, HBM, CoWoS packaging, specific materials. cyberagent operationalizes that thesis: it walks the supply chain down to the link "no amount of money can buy", and then applies hard anti-narrative discipline so it doesn't mistake a headline-driven spike for an opportunity.

It does not predict prices. It produces facts, a falsifiable logic chain, and monitorable physical signals — the final decision is yours.

---

Intellectual foundations

cyberagent stands on two ideas and turns them into a reproducible, falsifiable agent chain.

Leopold Aschenbrenner — Situational Awareness (the why)

In Situational Awareness: The Decade Ahead, Aschenbrenner argues that AI scaling is a massive industrial process, not a software one: every frontier model needs a bigger cluster, then bigger power plants, then bigger fabs. So the binding limits are physical — power, transformers, HBM, CoWoS packaging, specific materials — and effective compute compounds at roughly an order of magnitude (OOM) per year (GPT‑2 → GPT‑4 → ~2027 AGI). The people with situational awareness build conviction from the trendlines years before consensus prices them. cyberagent treats the market as that physical system; the thesis is distilled in references/sa-canon.md.

Serenity & Crux — the bottleneck method (the how)

The practitioner discipline comes from supply-chain bottleneck hunters such as Serenity (@aleabitoreddit) and Crux Capital. Instead of asking "which stock goes up?", they take the machine apart and look for the chokepoint:

What does the machine actually look like? Which part of its BOM is the hardest to replace? If one supplier stopped shipping tomorrow, how long would the whole chain wait?

• Serenity is narrow and deep — find one decisive choke point and concentrate.
• Crux is wide and disciplined — map a ~6-layer stack and size each layer by execution certainty, separating proven executors from early optionality.

cyberagent distills that discipline into a fixed, falsifiable chain that any LLM can run across any market.

We borrow their method — tracing a supply chain to its physical chokepoint, asking "where does the chain break", separating execution from optionality. We do not impersonate them, quote them, or present their positions as fact.

---

How it works

The chain is a telescope — it zooms from physical reality down to the specific, actionable name, one grounded LLM call per stage (each reads the upstream reports):

Positioning → Physical World → Human Development → Economics → Company Financials → Leaders & Verdict

That telescope is how the five-step method above (bottleneck → uniqueness → commercialization → elasticity → consensus) actually gets executed.

Phase 0 — Positioning. From the fundamentals, lock down what the company actually sells, then pin it to a specific layer of the physical / AI supply chain (materials → substrate → equipment → packaging → device → module → system → end demand) and a concrete machine (e.g. a GB300 NVL72 rack, a 1.6T optical link).

Five departments, run in sequence, each reading the upstream reports:

Dept	key	What it does
🪨 Physical World	physical	Locate the binding bottleneck on the SA ladder (power > CoWoS/HBM > raw logic); classify the asset as owner / adjacent / derivative / none. Non-owner ⇒ downgraded, scarcity-rent logic forbidden.
🌍 Human Development	human_dev	Place the demand on the AGI / OOM arc — early (runway left) or mature/peaked?
💱 Economics	economics	ore-seller vs processor; decompose the price move into earnings-growth vs multiple-expansion; detect valuation-framework switches; is it already priced (Gray Rhino vs loud consensus)?
📈 Company Financials	financials	Fundamentals + financial elasticity (linear vs non-linear); attribute earnings anomalies before flagging them.
🎯 Leaders & Verdict	leaders	Two-axis verdict — bottleneck identity (a) vs pricing position (b) — steelman + Munger inversion, monitorable exit signals, final decision.

The discipline (why it won't chase a top)

This is the part textbook frameworks skip:

• Real-time grounding — with Gemini it searches why a price moved (the catalyst, who said what), instead of trusting model memory.
• Price-action guardrail — the data layer flags parabolic / near-high moves; a stock that doubled in days on one headline is an AVOID / observe form, never a buy.
• Evidence ladder — every key claim is tagged Confirmed / Inferred / Weak; a load-bearing Inferred claim caps the confidence.
• Two independent axes — "is it a bottleneck" (classification) and "should you buy it here" (pricing) are never conflated. A non-bottleneck can be a fine trade at a price; a real bottleneck at a top can be a bad one.
• Honest "too late" — parabolic move + extreme valuation + loud consensus ⇒ the label is "too late / top", not an opportunity.

Educational and research use only. Output quality varies with the model, data, and many non-deterministic factors. This is not financial, investment, or trading advice.

---

Quickstart — 30 seconds

python3 -m venv .venv && source .venv/bin/activate   # isolated env (Win: .venv\Scripts\activate)
python3 -m pip install 'cyberagent[all]'             # everything: market data + all LLM providers + web UI
cyberagent                                            # launches the interactive wizard

No need to pick a provider at install time. cyberagent then walks you through it step by step — ① language → ② model → ③ paste your API key → ④ enter a symbol (NVDA / 600519 / 0700) — and prints the report. That's it. (Prefer the browser? cyberagent serve for the local web UI.)

Use python3 -m pip (not bare pip) and the venv above — on macOS a plain pip is often missing (command not found) and recent Python blocks installs into the system environment. The venv sidesteps both.

Want a leaner install? You don't need every provider — install just one with its extra (gemini / deepseek / openai / claude), e.g. python3 -m pip install 'cyberagent[stocks,web,deepseek]'. The wizard then offers that provider; for non-interactive use pass --llm <provider> (see below).

Use it from Python

import asyncio
from cyberagent import AnalystChain

chain = AnalystChain(llm="gemini", api_key="...", lang="en")
report = asyncio.run(chain.analyze("NVDA"))

print(report.final_decision)                   # ACCUMULATE / HOLD / REDUCE / AVOID
print(report.departments["leaders"].markdown)

(Inside Jupyter or an async app, await chain.analyze("NVDA") directly. Pick the report language with lang="en" / "zh" — the whole report is generated in it. Full API: docs/quickstart.md.)

More — other LLM providers · custom adapter · install options · CLI flags

Providers. Gemini is the default and the only one with real-time grounding; any of these works:

from cyberagent import AnalystChain, LLMAdapter, MockLLM

AnalystChain(llm="openai",   api_key="sk-...")
AnalystChain(llm="claude",   api_key="...")
AnalystChain(llm="deepseek", api_key="...")
AnalystChain(llm=MockLLM())                    # offline, no key — try the flow

class MyLLM(LLMAdapter):
    async def complete(self, system: str, user: str) -> str: ...
AnalystChain(llm=MyLLM())

Keys come from the argument, the environment, or a local .env (all variables: .env.example). Get one: Gemini (free) · OpenAI · Anthropic · DeepSeek.

Install options. Bare pip install cyberagent is the zero-dependency core. Extras: stocks (yfinance) · gemini / openai / claude / deepseek (providers) · web (local UI) · all (everything). DeepSeek is OpenAI-API compatible, so the deepseek extra is an alias for openai — install with python3 -m pip install 'cyberagent[deepseek]', set DEEPSEEK_API_KEY, run --llm deepseek.

CLI.

cyberagent analyze NVDA --llm gemini --lang en
cyberagent analyze AAPL --depts physical,economics,leaders   # subset, faster
cyberagent serve                              # local web UI at http://127.0.0.1:8000

The CLI and web UI auto-load .env and show a model picker (✓ next to every key found), live per-department progress, and the rendered report.

---

Use as a Claude Skill — no install

The whole methodology is also packaged as a self-contained Claude Skill in SKILL.md — it runs the same physical-bottleneck chain in pure-prompt form, no Python required. To install it in Claude Code:

mkdir -p ~/.claude/skills/physical-bottleneck-analyst
curl -fsSL https://raw.githubusercontent.com/CyberK13/cyberagent/main/SKILL.md \
  -o ~/.claude/skills/physical-bottleneck-analyst/SKILL.md

Then just ask: "analyze NVDA" — Claude picks the skill up automatically. (Any other agent that loads skills works the same way: give it SKILL.md.) The Python package adds live data, real-time grounding, and the CLI / web UI on top.

Methodology & prompts — fully open

There is no paywall. How to hunt a physical bottleneck is framework knowledge, not alpha. The complete system prompts live in src/cyberagent/prompts/departments.py, and the Situational Awareness anchor (the physical-bottleneck ladder + the OOM development arc) is distilled in references/sa-canon.md.

---

Roadmap

• LangChain / LangGraph tool wrapper
• MCP server (Claude / Cursor)
• EDGAR (US filings) + Tushare (A-share) adapters
• Segment-level chains for conglomerates
• Structured per-department gate verdicts (machine-enforced "stop")

---

Disclaimer

final_decision, confidence, and the department reports are AI-generated educational outputs, not financial advice. LLMs make mistakes; markets are unpredictable. Do your own research. The authors and contributors are not liable for any decision made based on this software. See docs/disclaimer.md.

License

MIT. See LICENSE.

Contact

Questions, ideas, or feedback? Reach out on X: @CyberK013.