The syllabus

• First 15 minutes: run a live agent. Trigger a live agent in the browser lab and see firsthand why a standard chat prompt breaks when real world inputs get messy.
• The anatomy of a RAC agent. Move from unstructured prompting to a structured plain text RAC definition. Map the agent’s Role, Business Context, Decision Logic, and Output Contract.

Lab: Take over the agent from the live build that is misclassifying edge cases. Fix its reasoning behavior by refining its plain text RAC rules.

• Plain text tool & MCP definition. Demystify external tool calling and the Model Context Protocol (MCP). Define which tools and data sources an agent can access, and the exact rules for calling them, in readable text.
• Memory & context window control. Define what an agent remembers across steps, what gets pruned, and how to prevent context bloat before it confuses the model.
• Multi agent handoffs. Wire specialist agents together by defining strict context handoffs in plain text.

Lab: Take the multi agent workflow from the live build and wire its tools, memory rules, and agent to agent handoffs in plain text so the team completes a multi step task end to end.

• Opening the black box. Read live execution traces in real time to diagnose why an agent made a bad decision instead of guessing.
• Infinite tool calling loops. Intentionally trigger and inspect an agent receiving ambiguous data and repeating the same tool call in circles.
• Context decay & instruction dropout. See how long workflows can cause an agent to forget its original rules mid task.
• Silent hallucinations & broken formats. Inspect what happens when an agent invents parameters or breaks the format expected by the next step.
• Cascading multi agent errors. Trace how one agent’s small mistake snowballs across an entire team of agents.

Lab: Inspect a broken multi agent workflow trace from the live build. Diagnose the exact failure pattern and stop the loop by adjusting the plain text definitions.

• Hard plain text guardrails. Move beyond polite prompt suggestions. Set maximum reasoning steps, blocked actions, and strict execution rules outside the prompt.
• Pass/repeat quality gates: LLM as a Judge. Build automated evaluator gates that inspect an agent’s output against your quality rubric and send bad work back for a retry before it moves forward.
• Human in the Loop systems. Design uncertainty triggers and approval gates so a workflow pauses and routes high stakes or ambiguous decisions directly to a human PM.
• Production codebase handoff. Package your tested plain text RAC definitions so your engineers or AI coding tools can implement the exact workflow directly into your company’s existing codebase, with zero framework lock in.

Lab: Add plain text guardrails, an LLM as a Judge pass or repeat gate, and a Human in the Loop escalation rule to the live build example. Practise the controls needed for production readiness.