AI agents, Claude and ChatGPT

Agentic trading: what it is, who offers it, and what can go wrong

In 2026, AI agents that plan and place trades under your rules moved from demos into brokerage accounts. What the term means, which brokers and exchanges offer it, what the SEC, FINRA, CFTC and ESMA have said, what can go wrong, and how to judge a platform before you connect money.

[

November 15, 2026

]
•
[

17

 

min to read

]

The short answer

Agentic trading means letting an AI agent plan and place trades for you within rules and limits you set. The agent reads market data, decides when your conditions are met and sends orders through a broker or exchange, usually from a separate account. Robinhood, Interactive Brokers and Coinbase all offer versions as of November 2026.

What is agentic trading?

Agentic trading, sometimes called agentic AI trading, is trading done by an AI agent that plans and acts for you inside limits you set. A chatbot answers your question and stops. An agent works toward a goal: it reads the market through tools, decides when your conditions are met, and places or proposes orders through a broker or an exchange. “Agentic” describes how much of that loop the software runs on its own.

Three things have to be true before trading is agentic:

  1. A goal in plain words. “Buy $200 of ETH if it closes the day under its 50-day average, with a stop 5% lower” is something an agent can act on. A hunch is not.
  2. Access. The agent can see prices and your account and can send orders, usually through an API or an MCP server. MCP, the Model Context Protocol, is an open standard that lets an AI app use another company’s tools, such as “get positions” or “place order.”
  3. Autonomy inside a box. It acts without you pressing every button, within walls you built: a separate account, size caps, stops and an approval setting.

The third point is what separates it from what came before. Algorithmic trading also runs on its own, but it follows code that someone wrote rule by rule. A chat assistant on its own can reason about a trade, but it waits for your next message and can’t reach your account. An agentic system sits in between: it takes instructions in words, like the assistant, and keeps working, like the algorithm. How an agent is built inside, from market data to guardrails, is covered in what is an AI trading agent?. This page is about the trend: the word, the products, the rules and the risks. AI trading as a whole covers more ground, from machine-learning signals to labeled bots.

Brokers spent 2026 building doors for agents, with dedicated accounts, MCP servers and approval switches.

Where does the word “agentic” come from?

“Agentic” is older than AI. Psychologists use it for agency, the capacity to act on purpose. In a 2001 paper subtitled “an agentic perspective”, Albert Bandura described human agency through intentionality and forethought, self-regulation and self-reflection. Swap the person for software and you have the modern sense: a system that sets its own steps toward a goal, acts, and checks its work.

In AI, the adjective spread with the agents themselves. On March 20, 2024, the AI researcher Andrew Ng began a series of letters in his newsletter, The Batch, under the title “Agentic Design Patterns”. He named four patterns: reflection (the model reviews its own work), tool use, planning, and several agents working together. He also gave the number that made the case: on a standard coding test, GPT-3.5 scored 48.1% answering in one pass and up to 95.1% when wrapped in an agent loop.

Brokers took up the word in 2026. On May 27, Robinhood announced that “the future of agentic finance has arrived” and called its new product Agentic Trading. Coinbase’s developer docs now describe “agentic accounts”, which “let an AI agent hold funds and act through a Coinbase account or wallet,” and Coinbase publishes open-source Agentic Wallet skills for AI agents. US regulators use other words again: the SEC’s exam staff and the CFTC now write “super intelligence (SI)” where most people would say AI, which helps when you search their sites.

What is Robinhood’s agentic trading, and how does it work?

Robinhood launched Agentic Trading on May 27, 2026, as a beta that supported stocks only at first. You open a dedicated agentic account, separate from the rest of your portfolio, and connect an AI agent you already use to Robinhood’s Trading MCP server. In Robinhood’s words, “your agent only has access to the funds you deposit into that account.” Every trade sends you a push notification, the app shows a live activity feed and the account’s profit and loss, and you can disconnect the agent with a tap.

Its help page for outside agents lists the assistants that work with the Trading MCP: Claude Code and Claude Desktop, ChatGPT, Codex, Cursor and Grok, plus Perplexity, Replit and others. The agent’s tools now cover stocks, options and crypto. On September 29, 2026, at its HOOD Summit, Robinhood added Robinhood Agents, an agent built into its app that runs on a model you choose, and said more than 150,000 customers had opened agentic trading accounts, with agents using its tools almost 30 million times a day. Loops, which will run a strategy on repeat around the clock, were announced as coming soon.

Trade approvals and their defaults, what the AI costs and where the product stops are covered in Robinhood Agents and the alternatives.

Which brokers offer agentic trading, and which let ChatGPT or Claude connect?

Robinhood is not alone. These are the brokers whose own pages, read on November 15, 2026, describe a way for AI agents to trade, with the assistants each one names. More may add MCP access at any time, so check the broker’s site before you rely on this table.

Brokers that let AI agents trade, and the assistants they name, November 15, 2026
PlatformHow your AI connectsAssistants it namesWhat the agent can tradeThe main limit
RobinhoodTrading MCP server, or Robinhood Agents inside its appClaude, ChatGPT, Codex, Cursor, Grok, Perplexity and othersStocks, options and cryptoA dedicated account holding only what you deposit
Interactive BrokersAI Integration, an MCP connectorChatGPT, Claude, Grok and Perplexity; Gemini coming soonStocks, options and ETFs on many marketsDrafts trade instructions for your review; Autonomous Trading only inside a funded AI Partition
CoinbaseCoinbase MCP (Coinbase for Agents), remote or as a command-line toolChatGPT, Claude, Claude Code, Codex CLI, Grok, Perplexity Computer and othersOver 900 crypto pairs, US equities and derivativesPortfolios you pick at sign-in; you choose autonomous or with your approval
PublicPublic MCP, or Agents inside its appChatGPT, Claude and moreStocks, options and cryptoEvery agent needs your sign-off before it runs
AlpacaAn open-source MCP server you run yourselfClaude Desktop, Claude Code, Cursor, VS Code, Gemini CLI and othersStocks, options and cryptoStarts in paper trading

Three details from those pages are worth reading twice. Interactive Brokers warns that if you link an AI any other way, “including browser-based and similar agents, you are providing those AI agents with full control of your account, without any controls or limits.” Its Autonomous Trading, introduced on October 21, 2026, lets a connected AI “submit orders without requiring approval for each trade,” but only inside a dedicated AI Partition funded from the main account. And Coinbase’s docs note that assistants behave differently with the same tools: “Claude previews and requests confirmation in the chat to create the order, whereas ChatGPT creates the order in one turn.” Before you connect anything, Coinbase suggests a separate portfolio funded “with only what you’re willing to risk,” to limit “the blast radius if the agent makes an unexpected trade.”

How people wire Claude to a broker, step by step, is in can Claude trade for you?. For ChatGPT, see can ChatGPT trade for you? For Grok, see can Grok trade for you?

What does agentic trading look like in crypto and on Hyperliquid?

Crypto had agent tools before the brokers did, for a plain reason: its markets never close and its accounts are wallets, which software can hold and sign with. Coinbase’s open-source agent toolkit has been on GitHub since late 2024, well before Robinhood’s launch. Three kinds of tools show what agentic trading looks like there in November 2026.

Wallets built for agents

That toolkit, AgentKit, is “Coinbase Developer Platform’s toolkit for giving AI agents a crypto wallet and onchain interactions.” It works with any AI framework and has an MCP extension. In February 2026, Coinbase published Agentic Wallet skills that let an agent sign in, check balances, send USDC and trade tokens on Base or Polygon from a command-line tool. Its Wallet MCP takes the opposite stance on control: “You approve every write action in Coinbase Wallet.” Coinbase also lets agents pay for data per request in USDC through a protocol called x402, so an agent can buy the research it trades on.

Agent keys on Hyperliquid

Hyperliquid, the onchain exchange best known for perpetual futures (contracts that track a price with no expiry date), has a name for this in its API. Its docs say API wallets “are also known as agent wallets”: your main account approves a separate key that signs trades for it, so software can trade without holding the account’s main key. Our Hyperliquid bot guide explains what such a key can and can’t do. Hyperliquid’s markets run around the clock and, through trade.xyz, include stocks, indices and commodities, so an agent there can act on Nvidia or gold on a Sunday (see how stocks trade on Hyperliquid).

Agents that run strategies

Products built on Hyperliquid turn plain-English strategies into running rules. Senpi, for example, describes itself as an AI trading agent on Hyperliquid with strategy templates and says its agent “has trade-only permissions”; it also connects to ChatGPT, Claude and Grok so they can “read your portfolio and place trades you approve.” Our Senpi alternatives page compares it with others. Ethy works on Hyperliquid too; its approach closes this page.

What do the SEC, FINRA, CFTC and ESMA say about AI agents in trading?

All four start from the same place: the rules they already enforce apply when firms use AI. Here is what each has published, newest first.

  • SEC, October 23, 2026. In its examination priorities for fiscal year 2027, the SEC’s exam staff says firms “are increasingly using super intelligence (SI) in core business operations,” investment analysis included, and that it will prioritize exams of how firms monitor SI, including in “trading functions.” For automated investment tools and trading algorithms, examiners check whether representations are “fair and accurate,” whether controls match what investors were told, and whether algorithms lead to advice that fits investors’ profiles. Like all staff statements, the priorities have no legal force of their own.
  • FINRA, December 9, 2025. FINRA’s 2026 Annual Regulatory Oversight Report added a section on generative AI that defines AI agents as “systems or programs that are capable of autonomously performing and completing tasks on behalf of a user.” It lists their risks: acting “without human validation and approval,” acting “beyond the user’s actual or intended scope and authority,” multi-step reasoning that is hard to trace or explain, mishandled sensitive data, too little domain knowledge, and badly designed reward functions. It asks firms to monitor what agents can access, keep a human in the loop, track agent actions and build guardrails. Its 2024 notice on AI adds that its rules are meant to be “technologically neutral.”
  • CFTC, December 5, 2024. A staff advisory reminded the firms the CFTC oversees of “their obligations under the Commodity Exchange Act and the CFTC’s regulations” as they begin to use AI. Its customer advisory on trading bots is blunter: the technology “can’t predict the future or sudden market changes,” and promises of enormous returns or 100% win rates are what scammers sell.
  • ESMA, May 30, 2024. The EU’s markets authority told investment firms that when they use AI, it expects them to comply with MiFID II, “particularly when it comes to organisational aspects, conduct of business, and their regulatory obligation to act in the best interest of the client.” It named algorithmic bias and poor data, opaque decisions, overreliance on AI by firms and clients, and privacy and security.

The common thread is responsibility. Regulators supervise the firms; the firms, in turn, put the trades on you. Interactive Brokers says “you are responsible for all orders submitted by your account,” and Coinbase’s agent docs say you are “solely responsible for reviewing and authorizing any trades, transfers, or account changes made through agentic workflows.”

What can go wrong when an AI agent trades with your money?

An agent can make every mistake a trader makes, plus a few of its own. Five are specific to agents.

1. Prompt injection

An agent reads text, and text can carry instructions. Coinbase’s AgentKit documentation puts it plainly: “LLMs do not reliably distinguish instructions from data, so text that reaches the model’s context can influence which actions the agent takes, including transfers.” It adds that this “cannot be resolved within the SDK,” and that AgentKit itself “does not gate transfers behind human approval, enforce spend caps, or allowlist destinations.” A social post, a web page or a token’s description can be written to steer an agent that reads it. The OWASP security project ranks prompt injection first in its 2025 list of risks for AI applications, and recommends giving a model the least access it needs and putting a human in the loop for privileged actions.

2. Loops that run away

Automation can repeat a mistake faster than anyone can stop it. The classic case involved no AI at all. On August 1, 2012, Knight Capital’s order router “rapidly sent more than 4 million orders into the market when attempting to fill just 212 customer orders,” in 45 minutes, and the firm lost more than $460 million, according to the SEC. An agent that retries a rejected order, misreads its own fill or reacts to its own trades can loop the same way. The fix is mechanical: caps on orders and position size that the platform enforces, and a kill switch.

3. The wrong order

Language is loose, and models fill the gaps. Coinbase’s docs warn that “lower-tier models may create orders with the wrong product, for example, ETH-USD instead of ETH-USDC,” and tell you to ask for approval before orders are created. A missing word can change the size or the side just as easily. A preview of the exact order, read before it goes out, catches most of these.

4. Limits that exist only in words

Telling an agent “never risk more than $500” is a hope, not a control. Coinbase says so outright about spending: “an instruction to the agent is not a server-enforced session spending limit.” Real limits live in the account: a separate pot of money, caps the platform enforces, and stops that sit on the exchange.

5. Nobody else to blame

FINRA’s list starts with agents acting without human approval and beyond the authority the user meant to give. When that happens with your money, the broker’s terms make the loss yours. And because “agentic” is a hot word, it is also a fresh label for old scams. A product that promises steady returns from an AI agent is the same red flag it always was, as we explain in is AI trading legit?.

How do you judge an agentic trading platform?

Ask these before you connect money. A platform you can trust answers each one in writing, and most red flags are visible from the outside.

Checklist for judging an agentic trading platform
QuestionA good answerA red flag
Where are the limits enforced?In the account: a separate account or portfolio, size and loss caps the platform enforces, stops resting on the exchange“Just tell the agent your limits”
What can it do without asking, and what is the default?A clear approval setting, shown at setup, that you can change at any timeAutonomy on by default, or no setting at all
What does it read, and can that text trigger a trade?Research kept apart from execution; orders come only from instructions or rules you approvedIt acts directly on posts, pages or messages it reads
Does it show the exact order first?A preview with the market, side, size, leverage and stopOrders go out from a vague request
What stops a loop?Caps on orders and position size, a pause after a run of losses, a kill switchNo ceiling on how often or how much it trades
Can you see what it did and why?A notification for every trade and a log with the reason for each actionOnly a profit and loss number
Can you test the exact rule first?A backtest of the same rule, or paper trading at live pricesReal money from day one
Who holds the keys?Trade-only permissions, and your login never handed to the AIA request for your password, seed phrase or private key
What does it promise?Tools, limits and a clear recordReturns, win rates or “the AI does it all”

If you want to weigh an agent against a plain rule-based bot task by task, trading bots vs AI agents does exactly that.

Where does Ethy fit in agentic trading?

Ethy is an AI trading agent you ask before you trade. It covers both halves of the loop, research and trading, and you choose how far it goes.

  • It reads and tests. It reads a chart on four timeframes together (daily, 4-hour, 1-hour and 15-minute), finds out why a market moved with a web search that cites its sources, and backtests a rule you describe in plain words on up to 365 days of history, with every entry and exit marked. It reads your trading plan, your own rules for risk, markets, process and goals, at the start of every chat, and says when an idea goes against it.
  • It trades when you ask. From the chat it places perpetual futures orders on Hyperliquid (crypto, plus stocks, indices, commodities and currencies on trade.xyz) and spot swaps on Base, X Layer and Robinhood Chain. New users get a confirmation card before each order; you can switch to “Execute without confirmation”. Either way, Ethy confirms in words anything you didn’t state, such as a stop it worked out for you.
  • It runs rules you activate. An automation trades on its own on Ethy’s servers once you press Activate, live or on paper. Automations can’t be triggered by news or sentiment.
  • Its limits live in code. Stop loss and take profit are native Hyperliquid orders that fire even when Ethy is closed. Automations add trailing stops, a maximum daily loss and a maximum drawdown in dollars, trading-hours windows, and a pause after 10 losing trades in a row, which is on by default. A riskier change to a live perp automation, such as a wider stop, needs your yes; every live trade sends you a notification; and archiving an automation can close its positions in one step.

Against the checklist above: the limits sit in the platform and on the exchange, not in a prompt; in chat, new users see each order before it goes out unless they switch that off; an automation trades only the rule you activated, which you can backtest or run on paper first; and every live trade is reported. Your balance lives in a wallet that belongs to you, and you can export its key at any time. What Ethy doesn’t do: it doesn’t trade options, and you can’t connect Claude or ChatGPT to it. Ethy is software for research and analysis; it doesn’t give investment advice.

Summary

  • Agentic trading means an AI agent that plans and places trades for you within rules and limits you set, through a broker’s or an exchange’s tools.
  • The word comes from psychology’s “agency”; Andrew Ng’s 2024 “Agentic Design Patterns” letters made it an AI term, and in 2026 brokers took it up, starting with Robinhood’s Agentic Trading on May 27.
  • As of November 15, 2026, Robinhood, Interactive Brokers, Coinbase, Public and Alpaca all let AI agents trade, and all of them name Claude or ChatGPT among the assistants that can connect.
  • In crypto, Coinbase offers AgentKit and Agentic Wallet tools, and Hyperliquid’s API has agent wallets that sign trades for a main account.
  • The SEC, FINRA, the CFTC and ESMA apply rules they already have; FINRA’s list of agent risks starts with acting without human approval and beyond the user’s intended authority.
  • The risks specific to agents are prompt injection, runaway loops, wrong orders and limits that exist only in words; judge a platform first by where its limits are enforced.

Questions people ask

Is agentic trading the same as algorithmic trading?

No. Algorithmic trading runs code that a person wrote rule by rule, and it does exactly what the code says. Agentic trading starts from instructions in plain words: an AI agent turns them into steps, reads data through tools and acts within limits. Many agents end up writing a fixed rule and running it, which combines the two.

What is Robinhood agentic trading?

It is Robinhood’s name for letting an AI agent trade for you. Since May 27, 2026, you can open a dedicated agentic account and connect an assistant such as Claude or ChatGPT through Robinhood’s Trading MCP server, and the agent can only use the money in that account. Since September 29, 2026, Robinhood Agents adds an agent built into the app.

Can ChatGPT or Claude trade stocks for me?

Yes, through a broker that offers an MCP connection. Robinhood, Interactive Brokers, Coinbase, Public and Alpaca all name Claude or ChatGPT among the assistants that can connect, each with its own limits, such as a separate account or an approval step. On their own, the assistants can research and draft orders but can’t reach your account.

Is agentic trading safe?

It is as safe as its limits. A separate account holding only the money you assign, caps the platform enforces, stops on the exchange and approvals while you learn make it manageable. Limits written only in a prompt are not limits: Coinbase’s own docs say an instruction to the agent is not a server-enforced spending limit.

What is an agentic wallet?

A crypto wallet that an AI agent can use within set permissions. Coinbase uses the name for open-source skills, published in February 2026, that let an agent sign in, check balances, send USDC and trade tokens on Base or Polygon. On Hyperliquid, an agent wallet is a key your main account approves to sign trades for it.

What do regulators say about AI trading agents?

FINRA defines AI agents as systems that perform tasks on a user’s behalf and lists their risks, from acting without human approval to going beyond the user’s intended authority. The SEC’s exam staff is reviewing how firms monitor AI, which it calls super intelligence, in trading. The CFTC and ESMA say their existing rules apply.

Sources

  1. Robinhood newsroom: Robinhood is Now Open to Agents (May 27, 2026)robinhood.com
  2. Robinhood newsroom: HOOD Summit 2026 (September 29, 2026)robinhood.com
  3. Robinhood help center: Onboarding an external agent (Trading MCP)robinhood.com
  4. Interactive Brokers: AI Integration and IBKR Autonomous Tradinginteractivebrokers.com
  5. Interactive Brokers: Interactive Brokers Introduces Autonomous Trading (October 21, 2026)interactivebrokers.com
  6. Coinbase Developer Docs: AI agents overview (Coinbase MCP, Wallet MCP, CDP MCP)docs.cdp.coinbase.com
  7. Coinbase Developer Docs: Coinbase MCPdocs.cdp.coinbase.com
  8. Coinbase Developer Docs: Agentic accountsdocs.cdp.coinbase.com
  9. Coinbase: AgentKit (GitHub, README and Managing Risk)github.com
  10. Coinbase: Agentic Wallet Skills (GitHub, February 2026)github.com
  11. Public: AI Agents and Public MCPpublic.com
  12. Alpaca: Trading API MCP serverdocs.alpaca.markets
  13. Hyperliquid Docs: Nonces and API wallets (agent wallets)hyperliquid.gitbook.io
  14. Senpi: home pagesenpi.ai
  15. SEC Division of Examinations: Fiscal Year 2027 Examination Priorities (October 23, 2026)sec.gov
  16. SEC: SEC Announces Fiscal Year 2027 Examination Priorities (press release, October 23, 2026)sec.gov
  17. FINRA: 2026 Annual Regulatory Oversight Report, GenAI (December 9, 2025)finra.org
  18. FINRA: Regulatory Notice 24-09 (June 27, 2024)finra.org
  19. CFTC: Staff Issues Advisory Related to the Use of Artificial Intelligence (December 5, 2024)cftc.gov
  20. CFTC: Customer advisory on AI trading botscftc.gov
  21. ESMA: ESMA provides guidance to firms using artificial intelligence in investment services (May 30, 2024)esma.europa.eu
  22. SEC: SEC Charges Knight Capital With Violations of Market Access Rule (October 16, 2013)sec.gov
  23. OWASP Top 10 for LLM Applications 2025: LLM01 Prompt Injectiongenai.owasp.org
  24. Andrew Ng, The Batch: Agentic Design Patterns Part 1 (March 20, 2024)deeplearning.ai
  25. Bandura, Social cognitive theory: an agentic perspective (Annual Review of Psychology, 2001)pubmed.ncbi.nlm.nih.gov

Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.

Ethy blog

Read more

The $ETHY ecosystem

Stop watching charts.
Start living.

Deploy your AI trading agent in one click.