Agentic trading is the practice of authorizing an AI agent — usually a large language model such as ChatGPT or Claude — to monitor markets and place trades on your brokerage account based on plain-language instructions you give it. The defining trait, and the reason the industry treats it as a new category, is that the agent takes actions. It does not merely answer questions like a chatbot; it watches conditions, makes decisions, and submits orders. Research firm Corporate Insight puts it plainly: the AI "acts autonomously on a user's behalf to monitor markets and place orders."
The term went mainstream in a six-week window. On May 27, 2026, Robinhood launched its Agentic Trading beta, letting customers connect third-party AI agents to a dedicated brokerage sub-account. On June 11, 2026, Coinbase followed with Coinbase for Agents, wiring ChatGPT and Claude directly into real customer accounts. A handful of brokers — Public.com, eToro, Gemini, Interactive Brokers, ThinkMarkets — had been shipping pieces of the same idea since 2025. So the question "what is agentic trading?" now has a concrete answer, a live product category, and an early, mixed evidence base.
The enabling technology is the Model Context Protocol (MCP) — an open standard that lets brokers plug their infrastructure into AI assistants people already use, rather than building proprietary bots from scratch. Robinhood's implementation is representative: users point their agent at a dedicated MCP endpoint, authenticate from desktop, and open a ring-fenced "agentic account." Supported platforms at launch included Claude, ChatGPT, Codex, Cursor, and Grok — essentially any MCP-compatible agent.
Once connected, the agent can read portfolio data, build and rebalance portfolios, run conditional strategies, and — where you authorize it — execute orders without per-trade confirmation. The guardrail pattern is consistent across the industry: a dedicated sub-account the agent cannot reach beyond, scoped permissions or API keys limiting what the agent may do, spending caps so it can only trade the balance you pre-load, a notification for every agent trade, and the ability to disconnect at any time.
Those guardrails limit the blast radius. They do not shift responsibility. Robinhood's disclosures are explicit:
"You are ultimately responsible for the trades your AI agent places in your account."
The firm disclaims responsibility for agent-generated losses, AI errors, and "unexpected agent behavior" — up to total loss of the funds you deposit into the agentic account. That posture is the industry norm, not the exception.
These four get conflated constantly, and the distinctions matter because each carries a different risk profile.
The shortest version: an algorithm follows rules it was given, an agent invents its own moves within the permissions it was given. We break this down at length in agentic trading vs. algorithmic trading, and compare the full spectrum in AI investing vs. algo trading vs. copy trading.
Robinhood launched Agentic Trading in beta on May 27, 2026 — third-party agents trading a dedicated sub-account via MCP, stocks and ETFs at launch, crypto added in early July 2026, access gated by email invitation. There is no separate fee beyond your own AI subscription. Our full Robinhood Agentic Trading review covers the setup, the guardrails, and the fine print.
Coinbase shipped Coinbase for Agents on June 11, 2026: ChatGPT and Claude connect to real Coinbase accounts for spot and derivatives trading, portfolio management, and USDC payments, with user-set spending limits, risk tolerance, and permitted trade types. Five days later it launched Coinbase Advisor, a conversational AI adviser operated by an SEC-registered investment adviser and CFTC-registered commodity trading advisor — one of the first fully registered AI advisers anywhere.
Public.com branded itself an "agentic brokerage" — the first to formally use the label — with its AI Agents release in March 2026: describe a strategy in plain English and the agent monitors conditions and executes autonomously across stocks, ETFs, options, and crypto. Desktop-only at launch.
eToro offers Agent Portfolios: users hand a separate sub-account to custom AI agents via scoped API keys, with a $200 minimum, combining scheduled rebalancing with signal-reactive trading. The company said usage of its automation features "nearly doubled" around the launch — a self-reported figure.
Gemini, the crypto exchange, launched its own Agentic Trading in April 2026, connecting third-party AI services with prebuilt "Trading Skills" modules.
Interactive Brokers took the conservative route: its Claude integration via MCP allows portfolio-specific analysis, and Claude can generate trade instructions — but a human must review and approve every order before submission. Nothing executes autonomously.
ThinkMarkets, a CFD broker, sits at the permissive end: its ChelseaAI lets users check positions, place orders, and move stop-losses through Claude, ChatGPT, or Grok via MCP, including direct order placement.
This is the question that matters, and the honest answer in mid-2026 is: the evidence does not yet show that LLM agents reliably make money.
The most instructive data point is Alpha Arena, a public experiment by Nof1.ai that gave six frontier LLMs $10,000 each of real money to trade crypto perpetuals with no human intervention. The spread of outcomes was enormous. Qwen3 Max finished Season 1 up roughly 22%. GPT-5 lost about 75% of its capital, running 17.2x leverage and adapting poorly. Gemini placed 238 trades and bled roughly $1,331 — about 13% of its entire stake — in fees alone. Claude went 100% long with no stops and was caught in a reversal. The real lesson: risk control and execution discipline beat raw prediction ability.
Academic work points the same direction. The STOCKBENCH benchmark found that most LLM agents "struggle to outperform simple buy-and-hold." The research literature is genuinely split: one ACM ICAIF 2025 study reported LLM-built portfolios beating benchmarks in-sample, while a 2026 arXiv study found LLM portfolios underperformed the S&P 500 on a risk-adjusted basis, with simple conversational prompts producing average monthly excess returns of 0.35% — statistically indistinguishable from zero. In-sample wins, out-of-sample shrugs.
None of that has slowed adoption. An April 2026 Investing.com survey of 938 U.S. investors found 62% now use AI to inform investment decisions, though 54% only "somewhat" trust the output and verify it elsewhere. Using AI to inform a decision and handing AI the account are very different commitments — and the data supporting the second one is thin.
Four risks deserve a paragraph before anyone connects an agent to real money.
Hallucination and misreads. LLMs fabricate, misread stale or incomplete data, and behave unpredictably. Robinhood's own risk disclosures concede the possibility of total loss of deposited funds from AI error alone.
No innate risk discipline. An LLM has no built-in concept of position sizing, drawdown limits, or fee drag unless those constraints are imposed on it — Alpha Arena's leverage blowup and fee bleed are the live demonstration.
Prompt injection. Because agents read the open internet, they can be manipulated through it. In May 2026, an attacker used a Morse-code-encoded tweet to prompt-inject a Grok-linked agent wallet and drain roughly $150,000 in tokens. This attack class is unique to agentic systems — a compiled algorithm cannot be talked into anything.
Liability sits with you. Platform disclosures uniformly assign responsibility for agent trades to the user, and regulators have noticed the gap: FINRA's 2026 Annual Regulatory Oversight Report flagged autonomous AI agents acting without human validation as an emerging risk, and in June 2026 House lawmakers formally demanded the SEC detail its oversight plans, with a response due July 31, 2026. As of this writing, no rules specific to agentic trading exist.
A disciplined checklist, drawn from how the platforms themselves structure their guardrails:
Agentic trading is real, shipping, and growing fast. What it is not — yet — is proven. The category is barely a year old, real-money results range from +22% to −75% across the best models available, and the one consistent finding is that discipline, not intelligence, separates the survivors.
No. Algorithmic trading executes pre-programmed, rule-based strategies that are written and tested before deployment. Agentic trading uses an LLM-powered agent that can reason, use tools, and adapt its own strategy in real time. An algorithm follows fixed rules; an agent improvises within permissions.
It carries meaningful, documented risks: LLM hallucination, absent risk discipline, prompt-injection attacks, and platform disclosures that place full liability on the user — up to total loss of funds in the agent's account. Platforms mitigate with dedicated sub-accounts, scoped permissions, and spend caps, but FINRA flagged autonomous AI agents as an emerging risk in its 2026 oversight report and no agentic-specific rules exist yet.
Robinhood (Agentic Trading beta, May 27, 2026), Coinbase (Coinbase for Agents, June 11, 2026), Public.com (AI Agents, March 2026), eToro (Agent Portfolios), Gemini (April 2026, crypto), and ThinkMarkets (ChelseaAI, CFDs). Interactive Brokers offers a Claude integration in which a human must approve every trade before submission.
There is no robust evidence yet that LLM agents reliably generate returns. In the real-money Alpha Arena experiment, results ranged from roughly +22% (Qwen3 Max) to −75% (GPT-5), and the academic STOCKBENCH benchmark found most LLM agents struggle to beat simple buy-and-hold. Winning results tend to come from backtests or in-sample studies rather than audited live trading.
No — that is the category's main pitch. You give instructions in plain English to an assistant like ChatGPT or Claude, and the broker's MCP integration handles execution. What you do need is a clear understanding of the permissions, caps, and risks you are accepting, because responsibility for every trade the agent places remains yours.