> ## Documentation Index > Fetch the complete documentation index at: https://docs.intention.xyz/llms.txt > Use this file to discover all available pages before exploring further. # AI trading: now and next > What already works when an AI agent trades on your behalf, what does not yet exist on any venue, and which parts of Intention are built for it. "AI-native" is a phrase that gets used loosely, so it is worth being exact about what it means here. It does not mean that agents can read this chain. Agents can read any chain. What is scarce is not market information — it is the ability to hand a real account to an AI agent, and still be the one who decides what that account can do. That capability is what this page is about: what exists today, where it is going, and which parts of the protocol are being built for it. ## Where trading actually is Three things are true at once, and they are at very different stages. **Most orders are already placed by code.** Programmatic execution, third-party front ends, market-making systems, and API clients account for the large majority of order flow on any modern venue. This is neither new nor controversial. The interesting question was never whether code would place orders — it has been placing them for years. What changes with AI agents is not that orders are automated, but who decides which order to place. **AI agents are starting to operate accounts rather than read charts.** Since late 2024, general-purpose models have been able to operate software the way a person does — reading a screen, clicking, typing. Through 2025, consumer brokerages shipped assistants that recognize charts and take spoken orders. In 2026, several major venues shipped agents that assemble a portfolio and place orders inside parameters the user sets. The direction is consistent across all of them: the model moved from advising the person to operating the account. **What has not moved is the account underneath.** Every one of those products runs on an account designed for a single human. The agent borrows that person's key. The books, the margin, and the risk parameters belong to the person, not to the agent. One authorization maps to one order, and everything else — what the agent may hold, how much room it has, what it may do next — lives outside the venue, in software somebody else operates. That arrangement works while a person is watching. It stops working when the entire point is that nobody is. ## The four phases | | Where the work sits | Where the venue sits | | ----------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------- | | **Phase 01 — Manual** | One person, one screen, one key. | The venue assumes someone is watching. | | **Phase 02 — AI-assisted** | The model proposes, the person confirms. | Unchanged. Today's agent trading products stop here. | | **Phase 03 — AI-autonomous** | The agent acts at machine speed inside the authority it was given. The person is out of the loop until something reaches a line they drew. | The venue has to hold the line, because nobody else is there to. | | **Phase 04 — AI-adversarial** | Agents trade against agents. | Market structure decides the outcome, not the interface. | The step from 02 to 03 is the hard one, and it is not a model problem. Models capable of Phase 03 already exist. What does not exist on any venue today is an account that can hold the authority: one that keeps its own books, reserves its own room to act, and can be called back in a single move. ## What that requires of a venue An account that can hold authority has to rest on a venue whose behavior can be checked rather than trusted, and that is not a feature a venue can add for AI agents. It is a property of where the trading logic lives. An AI agent needs to predict what an order will do before submitting it, which requires that the same input produce the same result on every node. It needs the price it reasoned about to be the price the position settles at, which requires the price to be certified in the block that consumes it. It needs the risk number it read to be the number the system acts on, which requires margin and liquidation to run inside the same execution step as the fill. And it needs its own history to be provable rather than reported, which requires every state change to carry the transaction that caused it. The [architecture](/protocol/architecture/overview) produces all of that, and it would produce it whether or not an AI agent ever connected. That is what makes it worth depending on: this is not a surface built for the use case, it is what a deterministic state machine over a consensus-committed ordering does. ## Where to go next The layers of the network, and how they produce the properties above. The execution layer these properties come from.