Path 07 of 07 · Developers and power users · advanced

From prompts to agents

By the end: You have one agent running against the AI API Gateway with least-privilege access and a confirmation step before any external action, built with Claude Code, Codex or a no-code agent designer.

The sequence

External Responsible Agentic AI The guardrails before the first agent: least privilege and read-only by default, separate service accounts, confirmation before external actions, dry-run modes, and no sensitive data in prompts, memory or tool descriptions. Open Guide Using AI Code Assistants With AI API Gateway | University IT Wire your tools to the gateway: Claude Code CLI and VS Code setup, Codex CLI, Cline, GitLab connection and UIT best practices. Assumes you already hold a gateway key from the Build with the API Gateway path. Open Course Agents and Workflows intermediate1h 30mOpenAI Academy OpenAI's advanced course on directing agent-assisted work: supplying context, defining outputs and boundaries, reviewing results and running a reusable workflow with human oversight. Free, with a certificate. Open Course Claude Code 101 intermediate1h 21mClaude Academy Build, debug and ship with Claude Code in the terminal, IDE and browser, from a first session to daily workflows. Point it at the gateway using the setup guide from step 2. Open Course Introduction to Model Context Protocol intermediate1h 4mClaude Academy MCP is how an agent reaches your own services: build MCP servers and clients in Python with tools, resources and prompts. Apply the least-privilege rules from step 1 to every tool you expose. Open KB article KB00020799: How to build a no-code agent in Gemini Enterprise Ship one: Agent Designer builds a scheduled Morning Email Digest agent from a plain-language description, with an Email connector, a preview run and a schedule. Needs the Gemini Enterprise Pro tier; agents are off on Standard. Open