learn-langchain-with-phoebe / Leader session 1 of 6
Learn LangChain with Phoebe · Leader track · Session 1 of 6

Chatbot, robot, or colleague? The autonomy ladder

"Agent" is the most-abused word in enterprise software right now. This session gives you the one mental model that cuts through it: a five-rung ladder from fixed rules to systems that choose their own steps. By the end you can place any vendor pitch, any internal proposal, and any headline on the right rung - and know what each rung costs.

🟢 Leader track Leaders: C-level · managers · curious minds No code, ever Start here
0-3 · Welcome 3-20 · Concepts 20-40 · Exercises 40-45 · Q&A
Part 0

Why this hour is worth your calendar

Agents are now a board topic. Budgets are being approved, vendors are pitching, and analysts are warning about cancellations - all using the same word to mean five different things. Gartner has a name for the worst of it: agent-washing, vendors relabeling ordinary chatbots and automation as "agents". Nobody on your data team can do that due diligence for you in a board meeting or a vendor call. That makes the vocabulary, and the skepticism, YOUR job. This track exists so you can do it in plain English, with zero code, ever.

Live - presented in session Self-study - read after class ★ Try it now prompt Sources covered at the end
★ What you walk out with today A five-rung autonomy ladder you can draw on a napkin, a two-question decision flow that sorts any use case onto the right rung, your own task portfolio actually sorted (that is the exercise), and a five-question screen that exposes agent-washing in any vendor meeting.
Part 1 · covers Anthropic's definitions + the agent-washing warning

Three machines that look alike 9 min live

A chatbot, an automated workflow, and an agent can produce identical-looking demos. Under the hood they are different machines with different costs, different failure modes, and different governance needs. The ladder below is the whole session on one picture.

Flexibility climbs with every rung ↗ Auditability and predictability fall ↘ Rules / RPA fixed scripts Single call one ask, one answer Workflow you fix the steps Agent model picks steps Multi-agent agents hand off Climb only when the rung below demonstrably fails you.
🔍 Click to zoom - the autonomy ladder: five rungs, one napkin drawing
LiveChatbot, workflow, or agent - definitions that hold up3 min

Anthropic's engineering team drew the line the industry now uses, and it fits in two sentences. A workflow is a fixed flowchart: YOU decided the steps in advance, and the model fills in text at some of them. An agent is a system where the MODEL chooses the steps - it decides what to look up, what to do next, and when it is done.

  • Chatbot: answers when spoken to. One question in, one answer out. Useful, cheap, and not an agent.
  • Workflow: "summarize the ticket, classify it, route it" - three fixed steps, forever, with a model inside some of them. Predictable and auditable.
  • Agent: "resolve this ticket" - the model decides whether to search the knowledge base, check the order history, escalate, or answer. Flexible, and much harder to predict.

Notice what the distinction is NOT about: model quality, vendor, or how impressive the demo looks. It is about one thing - who chooses the steps.

★ Try it now (any chat AI)Explain the difference between an AI workflow and an AI agent using Anthropic's definitions, in language a board member would enjoy. Then give me one example of each from a retail business, and tell me which one is harder to audit and why.
LiveRPA vs agents - and the 2026 hybrid pattern3 min

Your organization probably already runs automation: RPA - robotic process automation - the software robots that click through screens and move data between systems. RPA is a fixed script: brilliant at doing the exact same thing ten thousand times, brittle the moment anything changes. Agents are the opposite: they reason toward an outcome, so they handle variation - and introduce unpredictability.

  • RPA: deterministic, cheap per run, fully auditable, breaks on exceptions.
  • Agents: flexible on exceptions, costlier per run, harder to audit, can surprise you.
  • The 2026 hybrid pattern: keep RPA on the deterministic 70% of a process, and point agents only at the exception-heavy 30% that used to land in a human queue. Not either/or - a division of labor.
Real world

The claims team that stopped arguing. An insurance operations group (anonymized) spent a quarter debating "RPA vs agents" for claims intake. The answer was the split: RPA kept the standard claims it already handled flawlessly, and a pilot agent took only the malformed, multi-document exceptions. The debate was never RPA or agents - it was where to draw the line between them.

LiveAgent-washing - the due-diligence trap3 min

Gartner has warned loudly about agent-washing: vendors relabeling existing chatbots, RPA suites, and scripted assistants as "AI agents" because the word now opens budgets. The product did not change; the label did.

  • Why it matters to you: agent pricing on chatbot capability means you overpay for the label; agent risk controls on a chatbot means you over-govern it - and agent claims with no agent architecture mean the roadmap you were sold does not exist.
  • The one-question test: "does the model choose the steps at runtime, or did your engineers fix the flow in advance?" Vendors selling relabeled workflows struggle with that question in the room.
  • The trap: the demo cannot tell you. A polished fixed workflow demos exactly like an agent. Only the architecture answer - and a trace of a real run - can tell you. Exercise 2 gives you the full five-question screen.
Part 2 · covers the start-simple hierarchy + McKinsey's lessons

The only rule that saves money 8 min live

Every rung up the ladder buys flexibility and costs money, speed, and auditability. The organizations shipping successful AI in 2026 share one habit: they start at the bottom of the ladder and climb only under evidence. Here is the whole discipline as a two-question flow.

New AI use case Steps predictable, same shape every time? Workflow cheap, auditable High variance AND tools genuinely needed? Agent flexible, costly Single LLM call - start here yes no yes no Most ideas end here. That is success.
🔍 Click to zoom - the two-question decision flow that sorts any use case
LiveStart simple - the rule with a pedigree3 min

Anthropic's advice to its own customers is disarmingly modest: a single LLM call is usually enough. Summarize this, draft that, classify these - one ask, one answer, no ladder-climbing required. Their standing rule: add complexity only when it demonstrably improves outcomes.

  • "Demonstrably" is the load-bearing word. It means measured, not felt. A demo is not a demonstration of improved outcomes; a comparison against the simpler version is.
  • As a leader, this hands you a wonderfully polite challenge for any proposal: "what did the simpler version score, and by how much does this beat it?" If nobody built the simpler version, the conversation is already valuable.
The budget intuition Complexity is a loan against future operations. Every rung up the ladder is borrowed flexibility that operations, audit, and finance repay monthly. Approve the climb the way you would approve debt: against evidence of return.
LiveMcKinsey's two lessons from 50+ builds3 min

McKinsey teams distilled six lessons from more than fifty agentic builds. Two of them do most of the work for a leader, and both are pleasingly deflationary:

  • "It's not about the agent, it's about the workflow." The value never comes from the agent as an artifact - it comes from redesigning the business workflow around it. Teams that bolt an agent onto an unchanged process ship a demo; teams that redesign the process ship savings.
  • "Agents aren't always the answer." From the people building them for a living. Low-variance, standardized processes are better served by rules, RPA, or fixed workflows - lower cost, higher auditability, no surprises.

Together they hand you the questions for any internal pitch: which workflow is being redesigned, and why is an agent the right rung for it? A proposal that cannot answer both is a technology looking for a problem.

Real world

The pitch that answered both. A logistics team (anonymized) proposed an "exceptions agent" and opened with the workflow slide, not the agent slide: here is the process today, here is the redesigned process, here are the two steps where variance genuinely requires model-directed decisions. Approval took one meeting. The pitch before it - "we want to build an agent" - had taken four and gone nowhere.

Self-studyCost intuition - what each rung multiplies2 min read

You do not need to price tokens to have the right instinct. Each rung up the ladder multiplies three things at once - and the multiplication compounds:

RungWhat multipliesLeadership read
Single callBaseline: one model call per taskCheap enough to try on anything
WorkflowA few calls, fixed and known in advanceCosts are predictable line items
AgentUnknown number of calls + tool runs per taskBudget becomes a range, not a number
Multi-agentAgents calling agents; failure modes multiply tooNeeds spend caps and monitoring by design

Tokens, latency, and failure modes all climb together. The runaway-cost stories in session a3 are what happens when this table is discovered in production instead of in planning.

Exercise 1 of 2

Sort your own portfolio ★ 12 min · pen and paper

The ladder becomes yours the moment you sort real work onto it. No code - a pen, the decision flow above, and eight honest minutes.

List 8 real tasks from your organization that someone has proposed (or you suspect someone will propose) "putting AI on". Mix sizes: a report, a customer reply, a reconciliation, an analysis.

Run each through the two-question flow: steps predictable? → workflow. High variance AND tools needed? → agent. Neither? → single call. Write the rung next to each task.

Count the rungs. In most honest portfolios the majority land on single call or workflow - if all eight landed on "agent", re-run the first question with a colder eye.

Circle the one task you (or your organization) would have over-engineered - the workflow that was about to be funded as an agent. That circle is today's return on 45 minutes.

★ Your AI sorting partner (paste into any chat AI)You are my sorting partner for AI use cases. I will list 8 tasks from my organization. For each one, ask me the two questions from the autonomy ladder: (1) are the steps predictable and the same shape every time? (2) is there high variance AND a genuine need for tools like search or database access? Then place each task on one rung: single LLM call, workflow, or agent. Be stingy - recommend the simplest rung that works, and flag any task where I seem to be over-engineering. Here are my tasks: [list your 8 tasks]
Exercise 2 of 2

The agent-washing screen ★ 8 min · your next vendor meeting

Five questions that separate an agent from a relabeled chatbot - politely, in one meeting. Practice them out loud once; they land differently when they sound routine.

"Does the model choose the steps at runtime, or is the flow fixed in advance?" The definitional question. Hesitation here tells you most of what you need.

"What happens on an exception it has never seen?" Fixed scripts break or escalate; agents reason. The answer reveals the architecture.

"Show me a trace of one real production run - every step, every tool call." Real agent products produce traces on demand. "We'll follow up" is a data point.

"What limits what it can do and spend - and what happens at the limit?" A real agent vendor has thought hard about guardrails and caps, because their other customers made them.

"Which parts of the product are deterministic automation, and which are model-directed?" The honest answer is almost always "a mix" - and an honest mix beats a pure-agent claim every time.

Keep the tone curious, not prosecutorial These questions work best delivered with genuine interest. Vendors with real agents enjoy answering them; only the relabelers find them hostile.
Homework

Before session a2 ◐ 30 min total

★ Questions to ask your data team this week
  1. Of everything we currently call an "agent", where does each one actually sit on the autonomy ladder?
  2. Which of our processes are predictable enough that a workflow or RPA would be cheaper and more auditable than an agent?
  3. Where have we added agent complexity without evidence that it demonstrably improves outcomes over the simpler version?
  4. If a vendor pitched us "agents" tomorrow, who runs our due-diligence screen, and what is on it?
  5. What is the simplest version of our most ambitious AI idea - and have we tried it first?
Source material

Official sources covered

The leader track teaches from published engineering guidance, consulting research, and analyst warnings - no vendor marketing. This page covers:

Anthropic "Building effective agents"Part 1 · workflow vs agent definitions · Part 2 · the start-simple rule
McKinsey's six lessons from 50+ agentic buildsPart 2 covers the two leadership lessons; the remaining four surface in a3 and a6
Gartner's agent-washing warningPart 1 · the relabeling pattern · Exercise 2 · the five-question screen
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · Anthropic's line between a workflow and an agent is...

Who chooses the steps - that is the whole distinction, and it decides cost, auditability, and governance. Not model quality, not vendor, not the demo.

2 · A vendor demo looks impressively agentic. The single best due-diligence move is...

A polished fixed workflow demos exactly like an agent - only the architecture answer and a real trace can tell them apart. That is the heart of the agent-washing screen.

3 · Your team proposes a multi-agent system for a process with the same four steps every time. You should...

Predictable steps = workflow territory. McKinsey's builders say it themselves: agents aren't always the answer, and each rung up multiplies cost and failure modes.

Leader session 1 cheat sheet · pin this

The ladderRules/RPA → single call → workflow → agent → multi-agent. Flexibility up, auditability down.
Workflow vs agentYou fix the steps vs the model chooses them. The only line that matters - everything else is marketing.
The decision flowSteps predictable? → workflow. High variance + tools needed? → agent. Neither? → single call.
The only ruleStart with a single call. Add complexity only when it demonstrably improves outcomes - measured, not felt.
McKinsey's two lessonsIt's not about the agent, it's about the workflow · agents aren't always the answer.
RPA hybridRPA keeps the deterministic 70%; agents take the exception-heavy 30%. Division of labor, not replacement.
Agent-washing test"Does the model choose the steps at runtime?" + "show me a trace of one real run."
Cost intuitionEach rung multiplies tokens, latency, AND failure modes. Approve the climb like you approve debt.