Agentic AI Factories

Duration: 2 Days Skills Gained

By the end of this course, participants will be able to:

• Separate the model call from the rules that call it

• Build a loop that stops on success, on a budget, or on a repeated failure

• Draw the same job as a map of steps, with a branch and a human stop

• Block a write until a check passes

• Read a trace and say whether the fault is the prompt, the tool, or the map

Who Can Benefit

This course is designed for:

• Developers and data scientists building an internal agent

• Technical leads reviewing a design before it touches production data

Prerequisites

Participants should enter this course with:

• Comfort reading and editing Python, or enough fluency to follow a guided builder track

• Prior exposure to tool calling, in a project or in a prompting course

Organizational Objectives

This course assists organizations to:

• Stop one-off agent scripts from becoming unowned production paths

• Share one pattern for timeouts, retries, traces, and a human gate

• Bound the loop before anyone asks for a larger model

Software

A modern browser and an internet connection. The lab image includes Python and a graph library. Keys are provided. You do not need your own cloud account. A guided track names the same steps in a builder. The build track writes them in code.

Module 1 — The Rules Around the Model

By the end of this module, you will be able to point at what the harness owns.

• The harness owns the input contract, the tool list, the budget, and the trace

• A prompt asks. The harness decides whether the run may continue

• Stopping, and writing once, are harness jobs

• A factory is that harness, used on more than one job

Hands-on Lab: On a brittle script provided in class, label the missing controls: budget, stop, trace, and the write.

Module 2 — A Loop That Stops

By the end of this module, you will be able to wrap a tool-using task in a loop with a written stop reason.

• Try, check, stop. Where extra tries waste calls

• Stop on success, on a step cap, on a repeated tool error, or on a failed goal check

• Put a structured result back in, not a dump of free text

• Every exit writes why it stopped

Hands-on Lab: Wrap one task in a loop with a hard step cap. Every exit has a stop reason on the page.

Module 3 — A Map of Steps, Including a Person

By the end of this module, you will be able to replace a straight chain with a map that can branch.

• A straight chain fails when a step must choose

• Steps, the lines between them, and the condition on a line

• A human stop is a step on the map, not a comment in the prompt

• The build track writes the map in the lab library. The guided track names the same steps in the builder

Hands-on Lab: Rebuild the Day 1 loop as a map with one branch and one human stop.

Module 4 — A Gate Before a Write

By the end of this module, you will be able to block a side effect until a check passes, and to read the trace.

• A cheap check: the shape is right, a source is cited, or a test passed

• No write until the check passes

• On a failure, say which layer: prompt, tool, or map

• A second agent is another step on the same map, after this loop is reliable

Hands-on Lab: Add the check and a blocked write. Leave with the trace and a one-page factory card: steps, budgets, gates, and owner.