Claude Agent SDK & FastAPI Production AI Agents 2026

Build a real support agent in Python: your own tools, FastAPI, streaming, auth, budgets, evals, Docker and Cloud Run
What you'll learn

---

Run the Claude Agent SDK from Python, set the options that matter, keep sessions, and know what each run costs
Give an agent exactly the tools it needs, write your own in-process tools, and gate every change in code
Defend an agent against prompt injection that arrives through tool results, and prove it with tests
Put the agent behind FastAPI with streaming, one session per user, structured answers, memory and subagents
Deploy it with Docker and Cloud Run, with the API key in a secret and a user without privileges
Add authentication, per-user daily budgets, rate limits, timeouts and clean handling of model failures
Make it observable with structured logs, and gate changes with an eval suite wired into CI
Run agent jobs on a schedule that alerts you when they do not run, and compare the SDK with Claude Managed Agents
Requirements

---

Comfortable with Python (functions, async/await helps) and the command line
Python 3.12 or later and uv on Windows, macOS or Linux
Claude Code installed: the SDK runs its command-line tool under the hood
An Anthropic API key with a few dollars of credit (the runs in the course cost cents each)
Optional: Docker, a Google Cloud account for the Cloud Run lessons, and a GitHub account for CI and scheduled runs
Description

---

This course contains the use of artificial intelligence.
An agent that answers a question in a notebook is easy. An agent that real people call over HTTP, that spends money on every request, that reads data other people wrote, and that has to be right more often than not, is a different job. This course is about that job.
You build one project from start to finish: Deskmate, a support agent for a small product team. It starts as a knowledge base and six tickets, and ends as a service with its own tools, a web page, sessions and memory, users with budgets, rate limits, logs you can query, an eval suite in CI, a security check that runs on every start, and a morning digest on a schedule. It runs in a container on Cloud Run.
Every number in the course is real. Every run on screen was actually executed with the Claude Agent SDK and checked. When something failed, you see the failure and the fix: the budget fuse that returned an error after the result, the three-second timeout that took eight, the eval that passed a change that invented a feature, the scheduled job that never ran.
What you'll practice
The SDK itself: query(), the options that matter, sessions, and what every run costs
Your own tools in the same process, a gate in code on anything that changes data, and hooks
External MCP servers, and prompt injection through tool results, measured and blocked
FastAPI around the agent: one endpoint, streaming with server-sent events, one session per user
Structured output with Pydantic, memory that survives, subagents, and a minimal web page
Docker and Cloud Run with the key in a secret, never in the image
Tokens and daily budgets per user, rate limits, timeouts, and model failures as proper HTTP errors
Structured logs and traces that answer the questions you ask when something goes wrong
Evals that can say no, wired into CI, and a security review written as code
One new capability end to end, a scheduled job that tells you when it did not run, and the same job on Claude Managed Agents, compared on time, cost and trade-offs
Thirty lessons and twenty-nine downloads: each one is the project at that exact point, with its tests and the scripts used on screen, so you can run the same commands yourself.
How this course is made: the instructor designs and reviews the course. Every agent run shown was actually executed and checked, and the terminal output on screen comes from those runs. The narration uses an AI voice, the slide illustrations are AI-generated, and the instructor greeting that opens the first lesson is an AI avatar created from the instructor's own photo.
Who it's for: Python developers who want to put a Claude agent in front of real users, and teams deciding between building on the Agent SDK and using a hosted agent runtime.
Who this course is for

---

Python developers who want to ship a Claude agent to real users, not just run a demo
Backend engineers who already use FastAPI and want to add an agent the right way
Teams choosing between the Claude Agent SDK and a hosted agent runtime, with measurements to decide
Homepage

---