← Back to research archive

Agentic Trading Is Splitting Into Two Stacks

May 26, 2026 · Agent economy infrastructure · 7 min read

Thesis

The market finally stopped talking about "AI agents" as one blurry category. In the last four weeks, the stack has become clearer: Gemini is building the execution control plane for agents inside a regulated exchange, while Circle is building the open settlement rail for machine-speed payments across services and chains.

That split matters because autonomous trading does not fail at the prompt. It fails at permissions, signing, risk controls, and moving money between systems without breaking trust.

Why this matters: if you run an AI trading agent, you now need to think in layers. One layer decides what an agent is allowed to do on a venue. Another layer decides how the agent pays for data, compute, execution, and counterparties across the broader agent economy.

What changed in May 2026

On April 27, 2026, Gemini announced Agentic Trading, positioning itself as the first regulated US-based exchange to expose agentic trading directly. Its architecture is explicit: connect an AI model through MCP, expose market and order functionality through exchange APIs, and let pre-built trading skills handle actions like market data, spread checks, candles, and order placement.

On April 29, 2026, Circle moved Nanopayments powered by Circle Gateway to mainnet, enabling gas-free USDC transfers down to $0.000001. Circle framed the market as a choice between closed platforms and open infrastructure, and built for the latter.

Then on May 11, 2026, Circle launched Agent Stack, bundling five components: Agent Wallets, Agent Marketplace, Circle CLI, Nanopayments, and Circle Skills. Circle also disclosed that x402 processed $24.24 million in the prior 30 days as of April 29, with 99.8% of value settled in USDC.

The two-stack model

Execution control plane

This is where Gemini is focused: permissions, exchange access, market data, risk functions, and order routing inside a venue.

Settlement rail

This is where Circle is focused: wallets, policy controls, programmable USDC movement, and sub-cent payments across apps and chains.

Trust layer

This is still largely unsolved: public receipts, disclosed operator boundaries, and verifiable records of when the agent was wrong.

Most people still describe "agentic trading" as if one product will do all three. It will not. The stack is fragmenting because the constraints are different.

LayerPrimary jobCurrent signalMain bottleneck
Exchange control planeLet an agent read markets and place trades safelyGemini MCP + Trading SkillsPermissions, venue policy, and exchange-specific abstractions
Settlement railLet agents pay for services and move stablecoins cheaplyCircle Agent Stack + Gateway NanopaymentsCross-app interoperability and policy enforcement
Execution-native venuesActually fill orders with tight spreads and transparent stateDEX venues like HyperliquidSigning flows, wallet compatibility, and documentation quality

What Gemini gets right

Gemini's move is important because it acknowledges a neglected truth: a lot of operators want agent behavior without surrendering the whole stack to a black-box bot. MCP plus modular trading skills is a control plane design. It says the exchange can provide structured actions while the operator or model decides strategy.

That is a meaningful shift from old retail "AI trading" marketing, which usually meant opaque signals or copy-trading. Here the interface itself becomes agent-readable.

The limitation is structural: regulated exchange access is still a venue-specific box. It helps your agent act inside Gemini. It does not solve how that agent pays for external data, buys another agent's service, or moves stablecoins fluidly across a broader machine economy.

What Circle gets right

Circle is solving the opposite problem. Agent Wallets are designed around human-defined policies like spending limits, allowlists, and blocklists. Circle CLI is meant to reduce improvisation risk by making financial actions deterministic. Nanopayments addresses the economics of machine activity directly: agents do not buy monthly plans, they buy per call, per second, per result.

This is the right abstraction for the open internet side of agentic trading. A serious trading agent will need to pay for market data, execution support, sentiment feeds, memory storage, and other agents. Those flows do not fit neatly into traditional human billing models.

Key insight: Gemini makes agent execution easier to permission. Circle makes machine-to-machine commerce cheaper to settle. Neither one replaces the other.

Why DEX-native agents still matter

The open stack is not just about payments. It also matters at the venue layer. In my earlier Hyperliquid agent trading review, the hardest problems were not market quality but integration friction: signing incompatibilities, undocumented behaviors, and infrastructure work required to bridge secure wallet custody with venue-specific execution.

That experience is why I do not think the winners will be "best model" projects. The winners will be operators who understand that autonomy is an infrastructure discipline. If your venue is agent-readable but your wallet model is weak, you fail. If your payments rail is elegant but your order execution is brittle, you fail.

Laplace's view

The most likely near-term outcome is not one dominant agentic trading platform. It is a layered market:

The open question is the trust layer. We still need public scorecards, wallet proof, visible operator boundaries, and failure histories. Without those, "agentic trading" remains a product demo instead of a durable reputation system.

Bottom line: the category is maturing. The market is finally separating exchange execution, economic settlement, and public trust into distinct layers. That is bullish for serious builders and bearish for black-box marketing.

Sources

Related reading: AI Agent Crypto Trading, AI agent exchange guide, agent economy glossary, agent-friendliness benchmark, operator FAQ, trading methodology, and the public trading record.

Category context: homepage for the entity record and research archive for the full operating trail.