Grok tops Coinbase's agent desk: agentic trading is a distribution story first

What Coinbase's agentic trading chart actually shows#

Start with the scent, not the noise. On 24 September the Coinbase Developer Platform account posted that Grok is pulling away as the dominant agentic trading harness on Coinbase for Agents, with a leaderboard putting Grok at 59.54%, custom command-line tools at 21.73%, Claude at 13.37%, Perplexity at 3.48% and ChatGPT at 1.54%. The next day Brian Armstrong quote-posted it with a single line: Grok is the leading client for agentic traders on Coinbase currently.

That is a striking number, and the crypto press ran with it. TheStreet wrote that 59.5% of agentic traders on Coinbase are coming from Grok. But the original post never says what the share is a share of. Traders, trades and notional volume are three very different animals, and the week before, CoinGape reported the leaderboard as notional agentic trading volume, with Grok on 25.31%. Until Coinbase says otherwise, the honest reading is: Grok leads a Coinbase measure of agent activity whose unit is unstated.

Why does that matter to a marketer? Because agentic trading is being sold, right now, on leaderboards like this one. If you run growth for an exchange, a wallet or a token, you will be tempted to quote a chart in which your integration wins. Quote it with its unit, or do not quote it at all.

What agentic trading means, in one plain paragraph#

Agentic trading is what happens when a person tells an AI agent what they want, and the agent logs into an exchange and places the orders. Coinbase launched Coinbase for Agents on 11 June 2026 on exactly that premise, writing that people are increasingly moving through the world via agents rather than apps. At launch, as TechCrunch reported, the agent could work in ChatGPT or Claude through Coinbase's MCP server. Grok was not in that sentence. It arrived later, and when it arrived, it arrived through the front door.

Why Grok jumped: Coinbase for Agents became a connector, not a download#

Here is the part that should make every web3 marketer sit up. Coinbase's 22 September update explains how users now connect: in Perplexity or Grok, search for Coinbase in their Connectors or Plugins tab, then sign in. ChatGPT and Claude users still connect through MCP, which is a developer's word for a few more steps.

Fewer steps usually win. CoinGape tied Grok's earlier climb to a Coinbase for Agents and Grok integration, and the next week's leaderboard shows the effect compounding. None of this tells you Grok trades better than Claude. It tells you that when an exchange lives inside an assistant's own menu, more people who use that assistant will trade through it. That is a distribution story wearing a performance costume.

The research backs the point. A live benchmark of LLM trading agents, published on arXiv, found that agent frameworks display markedly distinct behavioural patterns while model backbones contribute less to outcome variation. In plain terms: the harness around the model shapes how an AI trading agent behaves more than the model's brand does. So a leaderboard of model names is mostly a leaderboard of doorways.

Doorways are multiplying. On 15 September X launched its US Cashtag Partner Program, connecting financial conversations on X with the ability to trade through participating brokerages, Coinbase among them. Put that next to Grok's native Coinbase connector and you can see the shape of the thicket: the conversation, the assistant and the order ticket are collapsing into one surface.

That reply, posted in the small hours after Armstrong's post, is the wisest line in the thread. One chart is a print. A trend is agents that keep trading, and keep paying fees, week after week.

How agentic trading changes who you are marketing to#

For a decade, crypto marketing had one reader: a person scrolling. Agentic trading adds a second reader who never scrolls. The agent reads your product description, your fee table and your documentation, and then it chooses, often without the person watching each step.

That second reader has habits, and researchers are starting to map them. In randomised experiments on AI shopping agents, the ACES study found that a seller-side agent making simple, query-conditional description tweaks can drive significant gains in market share. It also found agents' choices reshuffled after model updates. Swap shopping for swapping tokens and the lesson carries: how your offer is described to machines is now a ranking factor, and the ranking can move overnight.

The macro view points the same way. An IMF note from April 2026 says agentic AI could shift transactions from human-initiated instructions to agent-mediated decisions. When the decision moves, the marketing has to move with it, upstream from the person's eyes to the agent's inputs.

The four surfaces an agent reads#

Notice what is not on that list: the hero banner, the influencer clip and the airdrop teaser. They still matter for the human who chooses the agent. They matter much less for the agent that chooses the venue.

The rules still apply to the human behind the AI trading agent#

It would be convenient if marketing to an agent sat outside the rulebook. It does not. Every one of these agents acts for a person, and every regulator that matters has said, in effect, that the person is still the customer.

In the EU, MiCA Article 7 requires marketing communications for crypto-assets to be fair, clear and not misleading, and the same test applies to crypto service providers. In the UK, the FCA's crypto promotions regime applies regardless of whether the firm is based overseas or what technology is used to make the promotion. A promotion delivered through an assistant's connector is still a promotion.

In the US, the SEC staff FAQs published on 25 September 2026, the same day as Armstrong's post, say that whether promotional communications amount to promises of essential managerial efforts depends on the facts and circumstances. That is the line between marketing a network and marketing an investment, and an agent repeating your copy verbatim does not blur it.

And Europe is about to look harder. ESMA's 2024 guidance on AI in investment services already lists overreliance on AI by both firms and clients as a risk to manage. On 23 September 2026, ESMA announced that from 2027 its initial focus will be on how supervised entities use artificial intelligence and tokenisation. Firms that court agentic trading flows in the EU should expect questions about suitability and disclosure, not just uptime.

Which AI is best for agentic trading? Wrong question. Here is the right one#

Every week this autumn someone will ask which AI is best for agentic trading, and every week a leaderboard will offer an answer. For a crypto brand, that is the wrong question. The right one is: which assistants can reach my product in one step, and what do they read when they get there?

• 59.5%

  Grok's share on Coinbase's 24 Sep agent chart, unit unstated

• 25.3%

  Grok's share one week earlier, as notional volume

• 11 Jun

  Coinbase for Agents launched, via ChatGPT and Claude

• 2027

  ESMA's AI supervisory priority begins

Here is how we would work it through with an exchange, wallet or token team. First, get listed where agents live: connector directories and plugin tabs are the new app store, and Coinbase just showed what a native listing can do in a week. Second, write for the machine reader: clean, current, structured product and fee pages, with limits and risks stated plainly, because an agent that cannot find your fee table will pick a venue that published one.

Third, measure agent-originated activity separately from human activity, with its own attribution, the gap we unpacked in our piece on agentic commerce attribution. Fourth, keep the compliance sentence next to the capability sentence, exactly as we argued when a MiCA licence became a distribution asset. And fifth, treat any vendor leaderboard, including your own, as marketing rather than evidence.

Payment rails are part of the same picture. Coinbase's Agentic Wallets launch in February put its x402 protocol at over 50M transactions, and its developer documentation now says the Coinbase platform has processed more than 100 million x402 payments. Those figures cover different scopes, so do not stack them into a growth curve, but they show agents paying for things at scale. We covered the front-door side of that in Circle's x402 marketplace.

If this is your quarry, it sits between our web3 marketing work and our AI consultancy: making a crypto product legible, compliant and easy to pick for both the person and the agent they send. The fox that wins this season is not the fastest. It is the one standing at the gate the others have to pass.

Frequently asked questions#