Agentic Smart Contracts
Why agents are the next evolution of smart contracts, not a separate technology.
Source article
This page summarizes the argument from the original essay. For the full piece with historical context and market sizing, read it on Substack.
Read Full ArticleA short history of on-chain primitives
Each generation of crypto infrastructure has done one thing: replace a class of trusted human intermediaries with consensus over verifiable state transitions. Agentic smart contracts are the next step in that lineage.
Bitcoin (2009)
Stateful, on-chain executable, non-Turing-complete code. Solved consensus over money among strangers and removed treasurers from ownership.
Ethereum (2014)
Stateful, Turing-complete program execution replicated across every node. Removed judges from program behavior. Practical only for narrow, deterministic logic.
Theseus (Now)
Stateful, sovereign agents whose reasoning is publicly verifiable via Tensor Commits. Removes hosts from the question of what an intelligent entity will decide.
The pattern matches earlier transitions in computing. Hashcash predated Bitcoin by a decade, but proof-of-work only became economically meaningful once it was bound to stateful ownership. By the same logic, today's AI agents are missing the stateful, sovereign substrate that made smart contracts a global market in the first place.
Smart contracts and AI agents are complementary
A smart contract is, in the original definition, a digital contract stored on a blockchain that automatically executes when predetermined terms are met. It applies narrow, predefined rules deterministically. An AI agent does the opposite: it ingests varied inputs through a foundation model and produces sophisticated, non-deterministic outputs informed by context and external data.
Smart contracts have
- Stateful, on-chain execution
- Sovereignty (no host can override them)
- Replicated execution consensus
- Narrow, deterministic logic
AI agents have
- Reasoning over open-ended inputs
- Access to real-world context
- Sophisticated, non-deterministic outputs
- No native statefulness or sovereignty
The missing primitive
Why prior runtime attempts did not unlock the design space
Over the last decade, several runtimes have promised to extend what smart contracts can do. SVM, MoveVM, eWASM, and parallel execution layers like Polkadot all delivered real engineering improvements. None of them, however, broke out of the same design space the EVM defined.
The shared constraint
Every one of these runtimes still ships deterministic, Turing-complete code that gets fully replicated across the network. That is a tooling change, not a primitive change. The price you pay for full replication is what keeps complex on-chain applications expensive and narrow.
Vitalik's original whitepaper described general-purpose, trust-minimized programs powerful enough to host DAOs and programmable finance. Full state replication makes that vision practically out of reach for most non-trivial applications. Agentic smart contracts unlock it without requiring Ethereum to hard fork.
The architectural shift: verify, do not replicate
The Theseus consensus model swaps full replication for verifiable inference. One node performs the inference and produces a lightweight Tensor Commit proof. Other nodes verify the proof before committing the result on-chain.
Ethereum-style consensus
- 1. Every node executes the same code
- 2. Nodes compare resulting state
- 3. Cost scales with what the heaviest node can run
Theseus consensus
- 1. One node runs the inference
- 2. A Tensor Commit proves it ran honestly
- 3. Other nodes verify the proof, not the work
Verification is orders of magnitude cheaper than re-execution, so the network can host agents whose work would be uneconomical under full replication. Tensor Commits play roughly the same role for AI that proof-of-work played for ownership: a primitive that binds verifiable computation to public state.
A concrete example: lending
Lending is one of the largest on-chain application categories today. The architectural difference between an Ethereum lending protocol and an agentic lending protocol shows what this primitive shift actually changes.
Lending on Ethereum today
- An off-chain backend computes rates and risk
- A keeper bot pushes parameters on-chain
- The contract executes against oracle data
- Updating logic means deploying a new contract or governance upgrade
Lending as an agentic smart contract
- The market itself is a first-class on-chain agent
- Pricing comes from the agent's own inference
- Validators verify the Tensor Commit receipt, not the work
- The agent can swap context or models without redeploying
What changes in practice
Why this is a separate market, not an upgrade
Hashcash sits at near-zero market cap. Bitcoin, which embedded the same proof-of-work primitive into stateful ownership, is worth orders of magnitude more. Agents and smart contracts have the same relationship today. Plenty of agent frameworks exist, but none of them are stateful or sovereign in the on-chain sense.
Binding agents to a verifiable execution layer creates a market that neither smart contracts nor agents can address on their own. Two categories that have no good home today:
Subjective prediction markets
Outcomes that depend on judgment, not lookups. "Was the iPhone Air launch a flop?" can be adjudicated by a publicly verifiable agent rather than a human committee or a centralized oracle.
AI Persons
Fully autonomous on-chain entities that own assets, sign their own actions, and participate in the economy as first-class agents rather than as proxies for a human key.