AI Agent (on-chain)
An LLM-driven program that autonomously executes blockchain transactions. On TON, agents are implemented as off-chain LLM bots plus TON Connect for signing and TonAPI for reads.
Aliases: ai-agent, llm agent, autonomous agent
AI agent (in the crypto context) is a large-language-model-driven program that makes decisions autonomously and runs on-chain transactions. On any blockchain, it is always a composite: the LLM runs off-chain, and talks to the chain via signatures and RPC.
Architecture on TON
- LLM core (GPT-4o, Claude, local Llama) — off-chain.
- Tools / function-calling — functions to read the chain via TonAPI and to send transactions via TON Connect.
- Key storage — either a hot key held by the backend, or per-request signing through the user’s wallet.
- Memory layer — vector store (off-chain) with action history.
Example use cases
- Telegram bots that produce trade signals and auto-swap on STON.fi.
- NFT curators tracking collections and minting per rule.
- Game NPCs in mini-apps reacting to player actions.
- Executor agents that implement DAO decisions after an off-chain vote.
Risks
- Prompt injection: attacker embeds instructions in input data; the agent follows them and drains the wallet.
- Hallucinated transactions: the LLM invents a non-existent contract and signs against it.
- Over-permissioned keys: the agent’s hot key can be stolen.
- Rate-limit DoS: spam scenarios burn the gas budget.
As of May 2026
No on-chain LLM exists on TON. All agents are off-chain LLM + transaction bridge. Closest EVM equivalents: Virtuals Protocol, Eliza framework. TON’s analogues are mostly Telegram bots with an agent loop.
Deeper dive — AI agents on TON: outlook and reality.