AI Workflows

The Agentic Trading Stack

A visual map of the AI trading agent stack: intelligence, agency, execution, and compliance, with the companies Coinbase Ventures placed in each layer.

Starbelo map of the agentic trading stack, with intelligence, agency, execution and compliance as rows of name chips

On November 17, 2026, Coinbase Ventures published an X Article, "Beyond the bot: Mapping the agentic trading stack," written by Jonathan King. Jason Yanowitz of Blockworks quoted the piece and wrote: "The fastest growing use case for our data in 2027 will be trading agents."

The article describes agentic trading as an AI agent running part or all of the trading loop under authority you have delegated to it. In Coinbase Ventures' words: "You set a goal and guardrails, and the agent can research the market, build a strategy, size positions, place trades, and monitor risk within the authority you've delegated."

The map above is a Starbelo graphic. The company names and layer groups follow the landscape in that article. It is not a screenshot of their chart, and it is not exhaustive. Some names on the landscape are Coinbase Ventures portfolio companies, which the article marks in the text.

The four layers

Intelligence

The data and models an agent uses to form a view of the market.

Agency

The harnesses, terminals and copilots that turn a model into an agent that can keep working.

Execution

The brokerages and exchanges that turn a decision into a filled trade.

Compliance and observability

The trust rails that record what the agent saw, decided and did, so someone can audit it.

Top creator toolsAffiliate links

How people use this

The possibility is that you describe a goal and the limits you will live with, in plain language. An agent can then research the market, sketch a plan, size positions and watch risk, staying inside the authority you gave it. Some products on the agency layer offer a practice mode, so you can look at a simulated result before any live trade. You still set the mandate. You still decide whether real money is ever on the line.

The article's own example is a request like building a momentum-tilted mix across crypto and semiconductor stocks, with a drawdown cap and a weekly rebalance. Treat that as an illustration of the workflow, not as a strategy to copy.

Grok Bot, Muse and dots appear on the agency layer as general-purpose harnesses. If you are choosing a personal agent for content or side-hustle work, start with the 5 personal AI agents, compared. For how those agents search, reach accounts and pay, see the a16z agent infra map. A local option on Starbelo is Atomic Agent. For pay-per-call research tools an agent can call, see Monid.

Prompt
I want to explore a paper portfolio, not a live trade. Using only public information, draft a plan for a [style] mix across [markets]. Stay inside these guardrails: [max loss, assets I will not touch, venues I will allow]. First show me the research, the proposed sizes and the monitoring rules. Do not place any trade, and do not give me a recommendation to buy or sell, until I approve a written plan.
Prompt
I came across [tool or venue name]. Using the four layers of an agentic trading stack (intelligence, agency, execution, compliance and observability), tell me which layer it most likely sits in, what it appears to do, and what I should verify on the official site before I connect a wallet, a brokerage or any live capital. Do not recommend a trade.

Not financial advice. Trading is risky. Nothing on this page is a recommendation to buy, sell or hold any asset, or to use any named product with real capital. Check official terms, fees and eligibility yourself, and get advice from a qualified professional if you need it.

Intelligence, in more detail

Agents need clean, machine-readable market data: prices, order books, filings, news, sentiment and onchain activity, with timestamps and sources. The article argues that much of today's data is still priced for a human at a terminal, not for an agent that may query all day. x402 is named as an open protocol that lets an agent pay for the data it needs, when it needs it, with stablecoins.

On the landscape, the market-data group includes Bloomberg, Surf, x402, S&P Global, Nansen, Agentic.market, LSEG, Dune, Cambrian, Blockworks, Chainlink and DripStack. The article also names Pyth as an oracle network alongside Chainlink. Pyth is in the article, not on the landscape graphic.

Official pages linked in the article: Bloomberg, S&P Global, Nansen, Dune, Surf, Cambrian.

Market reasoning models are described as models trained to reason about financial markets, test an idea, and manage risk over a long unsupervised workflow. The landscape shows Prodigy Research, EdotEnv, Nof1, Duon Labs, EquiLibre and Grace IM. Model inference on the landscape is Venice, BlockRun, Surplus AI and OpenRouter. The article says inference platforms and routers like Venice let an agent pick a model for each task and, in some cases, pay per request in USDC.

Agency, in more detail

General-purpose harnesses give a model memory, scheduled tasks, tools, a sandbox and a way to recover from errors. On the landscape: Nous Hermes, OpenClaw, Muse, Instinct, Dots and Grok Bot. Instinct also appears again under trading agent harnesses on the same landscape.

Trading agent harnesses are built for the gap a general toolkit leaves: position sizing, take-profit and stop-loss limits, and a way to test a strategy before live capital is involved. Senpi is described as an "AI quant engineer" that turns a strategy written in plain English into code and can run it across Hyperliquid markets, with risk-based sizing and trailing stops. Supertake is described for everyday investors: you write an investment idea, or "take," and the agent turns it into a portfolio, with a practice mode and a public leaderboard. Also on this group: Listen, Instinct, Beep and Button.

Agentic trading terminals are full workspaces for research, strategy design, backtesting, execution and monitoring, often across stocks, crypto, perpetual futures and prediction markets. Many connect to existing brokerages rather than holding the assets themselves. Catalyst is described as turning a market view into a strategy across stocks, options and prediction markets. TrueNorth is described as giving people personalized agents that scan equities, commodities and crypto perp markets and set up trades. Also on this group: Minara, Astera, Questflow, Rockflow, Scalar Field, Stablejack, Cod3x and Velvet Capital.

Onchain trading copilots are chat-first assistants, often inside apps people already use, such as X, Telegram or a wallet. Funds usually stay in the user's own wallet. HeyElsa is described as helping people build a portfolio, swap across chains and earn yield through conversation. Bankr is described as letting people buy, sell and swap tokens, or set limit orders, by tagging it on X. Also on this group: Ask Gina, Milo, Dibs, Chance, Engine, Wayfinder, Glider and Weavr.

The article's view is that these product types are already blurring, and that the lasting question is who owns the decision loop.

Execution, in more detail

This layer routes orders, tracks status, reconciles positions and holds assets. The article describes a move toward one credential that can reach several asset classes.

Brokerages and exchanges on the landscape: Coinbase, Robinhood, Kraken, Gemini, Revolut, Binance, Composer, Public, Polymarket, Hyperliquid, Alpaca and Definitive. The Gemini in this group is the Gemini exchange, not Google's Gemini app.

The article names two ways venues are opening up to agents. One is bring-your-own-agent access, where an outside agent such as ChatGPT, Claude or Grok connects through an MCP server, a command-line tool, an API or a dedicated agent account. Coinbase for Agents is the example: a third-party agent can trade spot crypto, US futures and equities on a user's Coinbase account, within limits the user sets. The other is built-in intelligence inside the venue's own app. Coinbase Advisor is the example. Alpaca and Definitive (Flash) are named as execution rails an agent can use without building them from scratch.

Compliance and observability, in more detail

Today's rules assume a human made the call. When an agent decides, someone still needs to answer what it saw, what it decided, why, and who is accountable. This layer records inputs, reasoning, actions and trades so the behavior can be audited.

The article calls this one of the most underbuilt parts of the stack, and points to agent identity and attestations, tamper-proof decision logs, verifiable performance tracking and reputation systems. On the landscape: Alien, t54 Labs, Lemma and AgentOps.

Related: the 5 personal AI agents, compared, the a16z agent infra map, Grok, ChatGPT, Claude, Gemini, Atomic Agent, Monid, the news item.

Source: Coinbase Ventures, "Beyond the bot: Mapping the agentic trading stack," November 17, 2026, and Jason Yanowitz on X, November 17, 2026. Company names and layer groups on the map follow that landscape, which the original marks as not exhaustive. Layer explainers are written in our own words from the article. We did not add companies that were not on the graphic or in the article text.

Top creator toolsAffiliate links

The Starbelo 5 · weekly

Subscribe to the Starbelo newsletter

5 AI money wins in your inbox every week. Free, unsubscribe anytime.

Weekly AI wins. Unsubscribe anytime. See Privacy.

Keep Exploring

Advertise hereReach creators starting from zero.