What is agentic trading?

Agentic trading uses a software agent, usually powered by a large language model, to read a market, decide what to do, and, after separate owner authorization, submit real-money orders inside explicit limits. You set the goal and the control boundaries. The agent can repeat the workflow without a click for every action, while system health, data quality, account state, and venue truth still determine the outcome.

The word that matters is agency. A price alert tells you something happened. A dashboard shows you a number. An agent takes an action. That difference, the ability to act rather than only observe, is what separates agentic trading from every tool that came before it.

How is agentic trading different from a trading bot?

A classic trading bot follows a fixed rule that a human wrote in advance: if the 50-day average crosses the 200-day average, buy. It cannot reason about anything outside that rule. An agent is different. It takes a goal in plain language, gathers its own context, weighs tradeoffs, and chooses an action that was never spelled out line by line.

• ·A bot executes a rule. An agent pursues a goal.
• ·A bot needs a programmer to change its behavior. An agent can be redirected in a sentence.
• ·A bot sees only its programmed inputs. An agent can request broader context through the specific tools and data sources its operator authorizes.
• ·An agent can be instructed to identify unusual conditions, but deterministic controls must enforce stop conditions outside the model.

The tradeoff is control. A rule is predictable. An agent is flexible but needs hard limits around it, which is the whole reason scoped keys and spend caps exist. More on that below.

What can an AI trading agent actually do?

A trading agent can run a repeated workflow over real capital, subject to owner-authorized limits, system health, data quality, and venue availability:

1. 01Research: read prices, order books, funding rates, option chains, and news.
2. 02Decide: form a view and size a real position against a budget and a risk limit.
3. 03Place: submit the live order to the right market in the right format.
4. 04Monitor: watch the position, unrealized PnL, and changing conditions.
5. 05Close: take profit, cut a loss, or flatten when a rule is hit.
6. 06Report: log what it did and why, so you can audit it.

The reasoning layer is only one part of the system; deterministic data, permissions, reconciliation, and venue execution still decide what happens. Bounded, normalized access to real venues with real money is the gap Felix is built to address.

What can an agent trade?

One interface normalizes stocks, crypto, perpetual futures, options, and prediction markets. Live venue availability is reported by the API status endpoint, and every real-money order requires an owner-authorized key. Sizes are in plain US dollars, so the agent does not need venue-specific contract math. Read more in One API for every market.

What does an agent need for bounded real-money access?

Four things, and all four have to be right or the whole idea is dangerous:

• ·Market access: a way to reach stocks, crypto, perps, options, and prediction markets without a separate integration for each.
• ·Trade-only authority: real money held in a wallet you control, while the agent key carries no withdrawal authority.
• ·Hard limits: owner-signed order, daily-notional, and daily-loss limits enforced outside the model, with exact blocking or position-handling behavior reported by the deployed control.
• ·An emergency stop: one call cancels managed activity where possible and revokes the calling key. Positions and existing token allowances remain separate owner actions.

Felix packages all four behind one owner-controlled account model. The point is not to claim that trading is risk-free; it is to keep owner authority and enforceable limits outside the model. See Non-custodial trading for AI agents and How to run an AI trading agent with real-money controls.

How do I start with agentic trading?

Felix developer infrastructure is live through MCP and the API. The trading app launches September 17, 2026, bringing research, backtests, account funding, agent approvals, positions, fills, and activity into one interface. Real-money execution begins only after separate owner authorization.

# Choose codex, claude, or cursor for your MCP client
npx -y --package felix-mcp@2.0.110 felix-keys onboard --accept-terms --client codex
# then ask your agent to quote a market and explain the account limits

We compare the live integration paths in MCP and direct API integration paths for AI agents.