What Is Agentic Trading? A Plain-English Guide

· Lore

“AI trading” has meant a dozen different things over the years — signal bots, copy trading, a large language model asked to guess next week’s price. Agentic trading is a sharper idea, and it’s the one worth understanding, because it’s what actually changes how strategies get run.

This guide defines it in plain English, walks through why the hard problem isn’t the model but the keys, and shows what a safe setup looks like.

Agentic trading, defined

Agentic trading is the use of AI agents to generate, evolve, and operate trading strategies end to end — not as one signal feeding a rules engine, but as the operator running the whole loop.

The distinction matters. Most “AI-powered” trading tools work like this: a model outputs a score, and hardcoded logic decides what to do with it. The model is one input among many. In agentic trading, the agent is the operator. It forms a hypothesis, sizes a position, places the order, watches the fill, and revises — the way a human trader would, but continuously and without fatigue.

That’s a genuine capability shift. But it comes with a genuine risk: an agent acting on its own needs authority, and authority over money is dangerous by default.

Why custody is the hard problem

Here’s the scenario that keeps this from being simple. To trade, an agent needs to move funds. The naive way to give it that power is to hand it your private key. Now an autonomous piece of software — one that can hallucinate, get prompt-injected, or simply misjudge — has unlimited, irreversible control over your wallet.

That’s not automation. That’s an unsecured liability.

The real engineering problem in agentic trading is therefore separating what an agent may decide from what it may execute without you — and doing it without ever giving the agent (or the platform) custody of your keys.

This is exactly the wrong place to cut corners. A custodial “agent wallet” that holds your funds on your behalf reintroduces the counterparty risk crypto traders left centralized exchanges to escape. And most agent wallets that claim to be non-custodial still rebuild your key inside an enclave they operate — a single point they could, in principle, reconstruct.

How to run an agent safely: bounded autonomy

The workable answer is bounded autonomy — the agent gets real authority, but only inside an envelope you define and can revoke at any moment. At Lore this takes two concrete forms.

1. The agent is a co-signer, never the owner. Lore runs on a non-custodial MPC (multi-party computation) wallet. Signing authority is split — you and the agent each hold a share, and a transaction requires a threshold (for example, 2-of-2) to sign. No single party, including Lore, ever reconstructs the full key. The agent can act, but it never owns your funds, and you can pull its authority instantly.

2. Autonomy is a dial, not a switch. Lore exposes three modes:

  • Plan — the agent proposes every action and you approve each one. Maximum control; good for building trust in a new strategy.
  • Auto — the agent executes freely inside a policy envelope: position caps, allowed markets, maximum leverage, daily loss limits. Anything outside the envelope bounces back to you for approval.
  • Bypass — you’ve granted scoped, full trust for a defined mandate. The fastest, for strategies you’ve already validated.

The policy envelope is what makes “Auto” safe. Guardrails like a hard notional cap, a drawdown kill-switch, and allowlisted markets turn vague trust into enforceable rules an agent physically cannot exceed. If you’ve read our note on the Kelly criterion, this is the same lesson applied operationally: an agent that trusts its own edge too much will oversize — so you clamp it in policy, not in hope.

What an agent actually needs to be good

Bounded autonomy keeps an agent safe. Making it profitable is a separate discipline, and it’s where a lot of agentic trading falls apart. A few of the concepts that matter most:

  • Regime awareness. A strategy that prints in a trend can bleed in chop. Good agents classify the current market state and adapt — see regime detection.
  • Microstructure sense. Whether to post a passive order or cross the spread is a live decision with real cost, governed by things like adverse selection and order flow imbalance.
  • Honest backtests. An agent that rewrites its own strategies has to validate them without fooling itself. That’s why proper methods like purged cross-validation matter — an edge that survives only one lucky backtest path is noise wearing a Sharpe ratio.

We keep a growing library of these concepts, written for traders and agent operators, in the Lore Learn library.

Where Lore fits

Lore is building the agentic trading layer for people who want autonomy without surrendering control. The wallet has been shipping non-custodial agent signing since 2023, earned a $250K Optimism RetroPGF grant, secured roughly $50M in assets, and supports both EVM and Solana. The initial focus is Hyperliquid traders: connect, set a policy, pick or fine-tune a model, and deploy from Plan to Auto.

Agentic trading isn’t “let a bot run wild with your money.” Done right, it’s the opposite — you keep the keys, you set the rules, and the agent does the work inside them.

Want to run an agent on your terms? Get early access →