Factored
Advanced Agents & Safety
Design production-grade agents: the 12-Factor principles, the A2A protocol, asynchronous queues, reliability patterns, and access control enforced in code.
About This Course
Advanced Agents & Safety is the Week 9 course in the Jed-AI Academy Tier 3 track. It is for engineers who can already build a tool-using agent with LangGraph and MCP, and who now need that agent to survive production: long-running work, failing dependencies, other agents calling it, and users whose data it must not overreach. The course moves you from agents that work in a demo to multi-agent systems that are resumable, asynchronous, resilient, and bounded by authorization enforced in code.
Each section pairs an external resource with original deep dives. Where a strong free course exists, it opens the subsection: DeepLearning.AI's A2A: The Agent2Agent Protocol and AWS Skill Builder's Amazon Bedrock Getting Started. Where none exists, the opener is a primary source, such as the open-source 12-Factor Agents guide or Auth0's article on access control for AI agents. The deep dives then go past where those resources stop, check every technical claim against official specifications and documentation, and connect each topic to the rest of the Jed-AI curriculum.
What You Will Learn
The first half of the course covers how production agents are designed and how they talk to each other. You will work through all twelve factors of the 12-Factor Agents guide in order: treating tool calls as structured output, owning your prompts and context window, unifying execution and business state, launching, pausing, and resuming agents, contacting humans through tool calls, owning the control flow, compacting errors into the context, keeping agents small and focused, triggering them from any channel, and treating each step as a stateless reducer. You will then learn the Agent2Agent protocol: Agent Cards and discovery, messages, tasks and their lifecycle, artifacts, streaming and push notifications, how A2A complements MCP, and the authentication, authorization, signing, and webhook controls the A2A 1.0 specification defines.
The second half covers how those systems keep working and stay within bounds. You will design asynchronous agent APIs with the request-reply pattern, choose between queues and topics, run competing consumers with correct visibility timeouts, make every side effect idempotent, and handle poison messages with dead-letter queues and partial failures with sagas. You will add per-tool circuit breakers, layered timeouts based on the AWS Well-Architected Generative AI Lens, failover-aware model fallback, and a graceful degradation ladder. Finally, you will compare RBAC, ABAC, and ReBAC for agents, carry user identity through delegation with OAuth token exchange, classify every action as a read, a write, or a destructive action, and enforce policy checks, confirmation gates, and rate limits in code. A two-day mission brings all of it together in one production-grade multi-agent system.
Who This Course Is For
This course assumes you are comfortable with Python, HTTP APIs and JSON, tool calling with an LLM API, LangGraph basics, MCP fundamentals, and the OWASP risk vocabulary from Agent Security & Governance. It does not assume any background in message queues, distributed-systems failure patterns, the A2A protocol, or authorization models; each of those is taught before it is used or tested.
How the Course Is Structured
A Welcome section explains how the course works, what to set up, and how to get model access. Section 1 covers the 12-Factor Agents guide across three subsections, one for each group of four factors. Section 2 covers the A2A protocol, first its foundations and then its relationship to MCP and its security model. Section 3 covers asynchronous and queue-based architectures, opening with the Amazon Bedrock fundamentals course. Section 4 covers reliability, first at the tool layer and then when the model itself fails. Section 5 covers access control, first the authorization models and then safety boundaries enforced in code. Each of these five sections ends with a Section Check quiz that tests only what that section taught. The course closes with the Jed-AI Mission, a two-day build submitted for staff review.
Time Commitment and Completion
The Jed-AI calendar allocates about ten hours to this week. The A2A vendor course runs 1 hour 27 minutes of video and the Amazon Bedrock Getting Started course is listed at 1 hour, with extra time if you run their code and tutorial. The mission is scheduled separately as a two-day build. Progress is tracked automatically: pages complete when viewed, Section Checks when submitted, and the mission when your submission has been assessed by staff.