## Weekly signal

This week (May 11–19, 2026) the trading world moved from experiments to operational agentic systems: a regulated exchange documented an active developer push to open its APIs to autonomous agents, a mortgage-markets vendor reported the first live AI-executed trade, regulators said they are using AI to police prediction‑market abuse, and major fintech vendors moved to offer governed agent platforms for banks. These moves show the agentic stack (MCP/A2A, execution rails, governance, observability) arriving at production-grade scale—and the compliance and risk questions that follow.

## What changed

1) Gemini signaled production readiness for agentic trading tooling. In its Q1’26 investor presentation and related filings (May 14, 2026) Gemini documents a Developer Platform that unifies REST/WebSocket/FIX and Model Context Protocol (MCP) access and says it is opening these APIs so AI agents can autonomously monitor and execute trades through customer accounts. The filing frames agentic trading as part of its roadmap across crypto, predictions, and future asset classes.

2) Mortgage capital markets saw a first live AI execution. Mortgage Capital Trading (MCT) announced on May 13–14 that Atlas—its generative-AI advisor—completed a TBA trade on a live mortgage pipeline using a Trade Execution Agent operating within client-configured guardrails. MCT positions this as the first AI system that recommended and executed a trade in mortgage secondary markets.

3) U.S. regulator disclosure: CFTC using AI in surveillance. In an interview published May 15, CFTC Chair Michael Selig told WIRED the agency is using machine‑learning and third‑party analytics to screen prediction markets (Polymarket, Kalshi and similar platforms) for insider trading and manipulation and to prioritize investigations for human review. That signals active regulator attention to agent-enabled trading behaviors.

4) Banks and fintechs are productizing governed agent platforms. Fiserv launched "agentOS" on May 14 as a governed operating system and marketplace for agentic AI across banking workflows (risk, AML triage, payments and operations), built with OpenAI and AWS integrations—which creates an off‑the‑shelf path for banks to add agentic automation that could touch trading and treasury workflows.

Secondary technical development: an arXiv paper (submitted May 11, 2026) introduced an "Agent‑First Tool API" semantic pattern intended to make APIs easier and safer for autonomous agents to consume—practical for trading systems that need richer execution/verify/recover primitives than CRUD APIs provide.

## What to do with it

- If you run execution systems or marketplaces: treat agent access as a new client class. Add identity-bound credentials, explicit tool contracts (what an agent may and may not do), programmable budget/volume caps, and real‑time observability for agent sessions. Start with phased betas and rate‑limited sandboxes.

- If you build algo/quant systems: instrument agents for provenance (evidence chains, confidence scores) and safe fallbacks. Implement pre‑trade soft checks and human escalation triggers for exceptions. Use the Agent‑First API patterns for preview/verify/recover flows to avoid catastrophic automated executions.

- If you are in compliance/legal: expect active surveillance from regulators (CFTC) on prediction‑market and cross‑venue agent flows. Document governance, runbooks, and audit trails now; plan for production subpoenas and cross‑jurisdiction cooperation.

- If you are an infrastructure or DeFi operator: focus on on‑chain settlement and agent payment primitives (e.g., Pay.sh / onchain API payments) and verifiable agent identity. Prepare for questions about extraterritorial enforcement and custody of agent wallets.

- Short timeframe actions: run a 4–8 week internal pilot with a narrow, instrumented execution agent; add monitoring dashboards that surface anomalous fill patterns; and do a tabletop for a regulator inquiry scenario.