we opened this one by reading the industry’s promise back to itself.
open any keynote from the last year. any product launch. any linkedin post from anyone selling ai. you’ve heard this:
“the agentic workforce is here. hire your ai employees. agents that work while you sleep. don’t build software — deploy a team of agents. this is the year of the agent.”
that’s the promise. an autonomous workforce you can hire like people.
here’s what’s actually shipping under that word. a chatbot with a system prompt. a zapier flow with one ai step in the middle. an autocomplete that finishes your sentence. a search wrapper that calls itself a research agent. every one of those got relabeled an agent in the last twelve months, because “agent” is a budget line now and “chatbot” isn’t.
the bubble is the word
joel’s reframe of the title — in his own cadence on air — was sharper than the rehearsed line:
“the bubble is not really the agents bubble — is how much we’re stretching the word agents to cover other things that are not really agents.”
a word that covers everything from autocomplete to a system that can wire money can’t price risk. buyers pay agent prices for workflow capability. or they get real autonomy with no guardrails. both are the bubble.
what an agent actually is
if the word is the problem, the fix is a working definition. joel’s, from the show:
“in my mind, a real agent has some memory. some tools. access to a loop that allows it to run more itself. an llm obviously. guardrails. monitoring. observability. and the main thing — it has a goal. so it doesn’t have a fixed way to do things. that’s what makes it an agent.”
a workflow has fixed steps you wrote. an agent has a goal you gave it and decides the steps itself. that’s the line. workflows are great — most “agents” being sold today are workflows, and workflows are useful — they just aren’t agents.
“if a chatbot is just answering questions, it’s a glorified llm with access to a database of q and a. that can become an agent if you give it a goal — sell a ticket, increase employee happiness by 10%, whatever. then it has tools, guardrails, memory, a target. now it’s an agent.”
a real one, from the actual work
the example that landed cleanest was from joel’s own team:
“we have agents that do the monthly kubernetes upgrade. the workflow version would block on the first edge case. an agent that month said — somebody forgot to set pod affinity on this new application, the upgrade is going to stall. let me fix that first. then continue. nobody told it to do that. it knew.”
that’s the line between a workflow with an llm step and an agent. and most things being sold are the first one.
when the agent is good enough, you watch the agent
the most operator-tier moment in the episode:
“we have our observability tools actually validating to see if the agent is not hallucinating. because it’s so good now that you can’t tell. that’s our concern. make sure the data the agents are producing is actually backed by the history.”
when the agent is good enough that you can’t audit it by reading the output, you build a meta-agent to audit it. that’s the real ops layer for ai work, and almost nobody is selling it.
next
we keep saying somebody has to operate this stuff. whose job is it? the agents are spreading across every department. nobody owns operating them. that’s the real question next week — actually wait, that’s not the title. next week is the one we promised on day one.
episode five: why enterprise buys slower than you think.