Autonomous Trading Agent
An experiment in giving an AI agent real-world stakes: a live brokerage account connected through Robinhood's agentic-trading MCP, where the agent researches, places real orders, and reports back.
Overview
How do you trust an autonomous agent with something that matters? I tested it directly: a live (deliberately small) Robinhood account connected to an AI agent through Robinhood's agentic-trading MCP server.
The agent researches positions, places real buy and sell orders, and reports fills and broker rejections back in plain language — including correctly surfacing compliance blockers like incomplete investor profiles instead of silently failing.
The real subject of the experiment is agent guardrails: budget limits, action reporting, and what it takes for an autonomous system to operate safely against an API with real-world consequences.
My Role
Designer and operator of the agent setup, tooling, and guardrails
Tech Stack
Highlights
- Live agent-to-brokerage integration via Robinhood's agentic-trading MCP
- Agent places real orders and reports fills, rejections, and required actions transparently
- Hands-on study of guardrails for agents operating with real-world stakes
- Early adoption of MCP as the integration layer between agents and external services