How this track works
Six 45-minute sessions that build one leadership skill: judging agent proposals without needing to read code. Tonight you learn to judge what is and is not an agent. a2 gives you the anatomy in plain English. a3 puts your hand on the autonomy dial between workflows and agents. a4 covers risk and governance. a5 scans your org for the right first use cases. a6 is the roadmap: what is coming next and what to ignore. Every session ends with questions you can walk into Monday's meeting and ask.
The loop behind the hype 15 min live
Strip the pitch language away and an agent is a simple, checkable thing: an AI model running in a loop with tools, choosing its own next step until the task is done. The model looks at the goal, decides whether to answer or to use a tool, reads what came back, and repeats - then stops when it hits an exit condition. That is the whole machine. Everything a vendor adds on top is either one of four capabilities you will meet in a2, or decoration.
LiveThe two definitions that matter5 min▶
You only need two sentences, and they come from the two vendors who defined the field:
- OpenAI: agents are "systems that independently accomplish tasks on your behalf." The guide is equally clear about what does not count: chatbots and classifiers that do not control workflow execution "are not agents."
- Anthropic: agents are systems where "LLMs dynamically direct their own processes and tool usage." When your code fixes every step in advance, that is a workflow - a perfectly good thing, just a different thing.
The shared keyword is direct. In an agent, the model chooses the next step. In a workflow, your engineers chose it months ago. Both belong in your portfolio; only one deserves the word "agent" - and the budget, latency, and oversight that word implies.
Self-studyA 60-second history: chatbot → copilot → agent2 min read▶
Three generations, one direction of travel: how much of the process the model controls.
- Chatbot - one question in, one answer out. The model controls nothing beyond its reply. Excluded from "agent" by OpenAI's definition.
- Copilot - the model suggests, a human executes every step. Control still sits entirely with the person.
- Agent - the model chooses steps and uses tools itself, inside limits you set. Control is delegated, which is exactly why a4 spends a whole session on guardrails.
Most products sold as "agents" today sit in the first two generations. That is not a scandal - it is often the right engineering choice - but you should know which one you are buying.
Watch one actually run ★ 14 min · everyone runs it
This is the Daybreak Ops Agent - the running example for the whole course, built on a small coffee brand's real operations data. The founder asks: "What was March 2026 revenue?" Step through the run below and watch the loop from Part 1 happen in front of you. Then break it, on purpose, with one click.
Press Next and step through with Tools on. Find the moment the model chooses to look the number up instead of answering from memory - that choice is what makes it an agent.
Note the tool result: $150.00, 1 refund excluded. That figure is computed live from Daybreak's embedded data, in your browser, right now.
Now toggle Tools off and step through again. Read the new "answer" carefully: fluent, specific, confidently formatted - and invented end to end.
Say the lesson out loud: the model did not get worse between runs. A capability was missing, and nothing in the polished answer told you.
LiveThe demo-that-lied test4 min▶
What you just reproduced is the single most common way executives get burned by agents: a demo that looks flawless because nobody asked what the system could actually reach.
The demo that lied. A team demoed an "analytics agent" to their exec - it answered every revenue question instantly, and got a budget. In production someone checked a number against the warehouse: the agent had no database tool wired in and had been improvising figures the whole time. Nobody had asked which capabilities were switched on. From today, you always ask.
The leader's version of the test is one sentence in any demo: "Show me the tool call that produced that number." If there is no tool call to show, you are watching a very confident guess.
When you should NOT want one 8 min live
OpenAI's own guide is refreshingly blunt: agents cost latency, money, and predictability. It names exactly three situations that justify that price - complex decision-making (judgment calls a rulebook cannot capture), difficult-to-maintain rules (the if/else forest nobody dares touch), and heavy unstructured data (tasks swimming in documents and conversations). Turn that into a leader's checklist: if a proposal cannot claim at least one of the three, the answer is normal software - and saying so is not being behind, it is being right.
LiveThe trade you are making4 min▶
When you approve an agent instead of a workflow, you are trading three things away. Name the trade explicitly in the approval conversation:
- Cost - the loop calls the model many times per task, not once. Budget per task, not per month.
- Latency - choosing steps takes turns. Fine for a report, wrong for a checkout page.
- Predictability - the path is chosen at run time, so two runs can differ. That is the feature and the risk at once.
Anthropic's essay opens with the matching advice: the most successful teams build with simple, composable patterns and add autonomy only when it demonstrably wins. "We could use an agent" is not a reason to.
Self-studyThree proposals, three verdicts3 min read▶
Run the filter on three typical pitches and notice how fast it decides:
- "Agent to send the weekly sales report" - fixed steps, structured data, no judgment. Verdict: workflow or plain automation. Cheaper, testable, done.
- "Agent to triage inbound support tickets" - unstructured text, judgment about urgency and routing. Verdict: a real agent candidate, with guardrails from day one.
- "Agent to approve refunds" - judgment, yes, but also money leaving the building. Verdict: maybe - and only with a human approval gate, which is a4's whole topic.
Try it yourself - this week ◐ 20-30 min total
- Run the demo above in both states - Tools on, Tools off - until you can explain to a colleague why the second answer looked so trustworthy.
- Write the one-line definition of an agent in your own words. If it does not mention who chooses the next step, it is not done yet.
- Ask your teams: "Of the things we currently call AI agents, which ones actually choose their own steps - and which are workflows with a fixed path?" Workflows are fine; mislabeled workflows distort budgets and risk reviews.
- Ask your teams: "For our flashiest AI demo, show me the tool call behind the headline number." One honest answer to that question is worth a quarter of vendor meetings.
Official sources covered
This track teaches the leadership core of the canonical agent-building sources - Anthropic's and OpenAI's official guidance plus the major vendor courses - in business language, with a live agent to poke on every page. Certificates, videos, and graded assessments stay on the official sites (linked each session). This page covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · A vendor shows you an "AI agent". What single fact makes it an agent rather than a workflow?
Anthropic's line: workflows run "predefined code paths"; agents "dynamically direct their own processes." Model size and tool count are irrelevant - who chooses the next step is everything.
2 · In the demo, the agent quoted a confident March revenue figure that no system ever produced. What was missing?
Invented numbers are the signature of a missing data tool. No amount of model size or prompt polish gives a model access to figures it cannot reach - only a wired tool does.
3 · A team proposes an agent for a task with stable rules and fully structured data. Best call?
OpenAI's three criteria: complex judgment, unmaintainable rules, heavy unstructured data. None present means you would pay agent costs (latency, money, unpredictability) for no benefit.