AI's Place in Crypto OS
Five roles, six capability levels, and one AI Safety Stack. Models propose; deterministic systems validate.
AI is not a chapter added at the end of this course. It appears in every chapter, always in the same role:
AI proposes hypotheses. You verify the facts.
This page explains what that means in different settings.
Five roles
“Using AI” can mean different roles with entirely different risks.
Explain complex concepts and ask follow-up questions until you understand them.
Search, organize, compare, extract, verify, and monitor.
Write PRDs, code, tests, debugging notes, audits, and documentation.
Continuously monitor protocols, wallets, markets, risk, governance, and news.
With identity, a wallet, budget, strategy, and execution ability, it becomes an economic agent.
Mistakes in the first three roles cost time. Mistakes in the last two cost money. From the fourth role onward, the course therefore requires accompanying controls.
Six capability levels
| Level | What AI can do | Required guardrails |
|---|---|---|
| L0 Do not use AI | Learn the fundamentals yourself. | — |
| L1 AI Assistant | Questions, answers, and explanations. | Treat conclusions as hypotheses, not facts. |
| L2 AI Copilot | Research, writing, and code. | Verify against primary sources; review the code yourself. |
| L3 Tool Agent | Call search, APIs, RPC, databases, and read-only wallets. | Allowlist tool permissions and log every call. |
| L4 Financial Agent | Analyze, trade, pay, and manage assets. | Budget cap, address allowlist, simulation first, manual approval above the limit. |
| L5 Autonomous Economic Agent | Earn income, buy services, pay costs, and cooperate with other agents autonomously. | Policy engine + Risk engine + Auditable logs, with mandatory one-click shutdown. |
A useful rule: your level depends on how much you can verify. If you cannot tell whether the model's contract call is correct, you should not operate above Level 3. You would be unable to detect a mistake.
AI Safety Stack
Every AI system handling funds must pass through this chain. It is the core of T32 and the acceptance standard for all agent assignments.
Three responsibilities summarize the chain:
- Model: Proposal.
- Deterministic system: Validation.
- Execution system: Execution.
Keep them separate. Once model output can trigger execution directly, all intermediate checks lose their purpose.
An architecture that is strictly forbidden
Architecture prohibited in this course
- Prompt
- LLM
- Private Key
- Send Transaction
The model must never receive a private key and send transactions without checks. This structure fails every assignment.
This structure has nowhere to detect mistakes. One hallucination, prompt injection, or incorrect unit can lose the funds, and afterward even the cause may be impossible to establish.
Any course assignment using this structure fails immediately.
How to read each chapter's AI Lab
Every AI lab has two parts: what AI should do, and the checklist you must verify yourself.
The checklist is deliberate. It addresses four common model errors:
| Error type | What it looks like | How to check |
|---|---|---|
| Invented facts | Nonexistent protocol parameters, interfaces, or reports | Open primary sources and verify each item |
| Outdated information | Changed parameters, deprecated versions, or protocols already attacked | Check documentation dates and current on-chain state |
| Incorrect definitions | Counting incentives as revenue, FDV as market cap, or TVL as revenue | Recalculate the key figures yourself |
| Correlation presented as causation | “A causes B,” when they merely occur together | Seek counterexamples: did A occur without B? |
After a dozen labs, you will develop a sense of where models are reliable and where they need close attention. That judgment is itself the AI literacy the course aims to develop.
Common misconceptions
Controls exist because mistakes in fund operations are irreversible.
On-chain transactions have no undo button. Even an error rate of one in a thousand eventually produces a failure in an automated system, and that one failure may cost all the principal. The issue is irreversibility.
Human approval works only when the person understands what they are approving.
A system that requests approval 200 times a day effectively has no approval: people start confirming without thinking. A policy engine should first automatically allow or reject most requests and send only decisions requiring judgment to a person.
They do not necessarily. Chapter 33 examines this carefully.
Traditional payment channels are usually more suitable when parties trust each other, amounts are large, and frequency is low. On-chain payments have a real advantage for tiny amounts, very high frequency, unfamiliar counterparties, and programmable settlement. Evaluate the setting first, then choose the method.
That leaves you with a search engine that speaks confidently.
The course aims to teach you to recognize when AI is talking nonsense. That judgment comes only from repeatedly verifying its work yourself.
Related chapters
| Topic | Part A | Part B |
|---|---|---|
| AI changes research | Chapter 31 | T31 |
| AI in trading and asset management | Chapter 32 | T35 |
| Why agents need Crypto | Chapter 33 | T33 |
| Agent wallets and permissions | Chapter 34 | T34 |
| Machine payment protocols | Chapter 35 | T34 |
| Reliable AI architecture | — | T32 |
| Autonomous on-chain economy | Chapter 36 | T36 |