The setup
Last week HubSpot held its Fall '26 conference in Boston. It's been called INBOUND for fifteen years. This year they retired the name entirely and rebranded the whole thing UNBOUND, then spent the keynote on something called "Growth Context."
The press mostly covered one thing: Breeze Assistant got promoted from question-answerer to orchestrator. You describe an outcome, and it picks which of HubSpot's eight agents — Campaign, Nurture, Deal Progression, Revenue, Customer, Prospecting, Data, Content — should go do it.
HubSpot's Director of AI Products, George Davis, was startlingly honest about why: "Most GTM teams are now running two or three agents, none talking to each other, all requiring active management. Coordinating AI has become its own job."
So Breeze routes. You don't manage the agents — you're managed past them.
Here's what the keynote didn't linger on: every one of those agent actions runs down a credit meter. Roughly fifty cents per "resolved" support conversation. About a dollar per recommended lead. The analyst's line: software "is being repriced from seats to outcomes."
Interesting, but not the part I care about.
The wall and the door
The part I care about is what HubSpot shipped right next to the orchestrator: the Agent CLI.
It's a command line built not for humans but for AI coding agents — Claude Code, Claude Cowork, OpenAI Codex — to read and write the CRM directly. It ships with every plan. It uses zero HubSpot credits. It's a separate binary from the developer CLI, and it ships with a --dry-run flag and JSONL in and out, so an agent can plan, preview, and apply changes the way a careful ops person would — except unattended, in bulk, at 2 AM.
HubSpot's CPTO, Duncan Lennox, said the quiet part out loud: "Agents don't click through dashboards or navigate interfaces; they call APIs, read structured outputs, and take action."
Hold those two next to each other. One product line says: run your agents inside our walls, we'll route them for you, and you'll pay us per action. The other says: bring your own agent, run it on your own infrastructure, reach into the CRM for free — you just have to own the agent.
A company named its conference "Unbound" and shipped, in the same release, the wall and the door.
The sharpest line I've read this week: "If you want to know what any software company really thinks about your freedom, skip the keynote and read the invoice."
The part that's mine
I run sales ops. HubSpot is my CRM. I'm the person in my company who's going to get asked, probably this quarter, "should we turn on the Breeze agents?"
And here's the uncomfortable part: I'm also the one person in the building who's been living on the free side of that split for a year.
I run four to ten Claude Code sessions a day. My second brain — GTBrain — is operated by an agent that reads my work and writes my notes. My coffee marketplace runs an eight-agent content pipeline. My entire night life is conducted through agent CLIs on boxes I own.
Last week, my day job's CRM shipped a command line built for exactly the tool I'm already fluent in.
Let me say it again, because it's the whole point. The interface I already live in — a terminal, an agent, --dry-run before you write — just became the interface of the system my paycheck runs on. The engineers at my company know Claude Code but not the CRM. The ops analysts know the CRM's UI but not Claude Code. I'm the only person who lives in both worlds.
That's not a brag. It's a map. It points at a question I've been writing around all month without knowing the vendor would hand me the answer.
What I'm doing about it
Two months ago I wrote a note to myself that's been sitting there, stalled. It listed six sales-ops decisions that are secretly optimization problems — territory design, lead scoring, routing, comp, SDR:BDR ratio, pipeline SLAs — and it ended on the exact bottleneck:
"Changing a territory in HubSpot means updating 30+ properties, rerouting active leads, adjusting quotas… An experimental loop requires programmatic CRM mutation — which is terrifying if done wrong, transformative if done right."
Terrifying if done wrong. That's the phrase that's been blocking me. The reason I never built the lead-scoring experiment harness or the territory simulation is that writing directly to the CRM from a script is how you break the thing everyone in the company stares at.
The Agent CLI ships with --dry-run. It acts as you, with your permissions. An agent plans and previews, a human approves, and then it applies. That's the entire safety story I've been missing — wrapped in the one interface I already know how to drive.
So this week I'm installing it. Read-only first. I'll run HubSpot's own sample prompts — the Monday-morning enrichment-gap report, the pipeline-hygiene check — and see whether my existing agent, on my own box, produces the report it currently takes an analyst hours to assemble. Without touching a HubSpot credit.
If it works, the argument I've been making all month finally gets its day-job proof. If it doesn't, the failure mode is itself the finding.
The actual lesson
All month I've been writing some version of the same sentence: own the layer, don't rent it.
The comp plan is the layer that matters, so version it. Lead provenance is the layer that matters, so make the lead prove it's real. The CRM is a cache, not a source of truth, so own the fact layer outside the vendor.
The Agent CLI is the next sentence in that argument, aimed at the interface instead of the data. Do you run your agents inside the vendor's walls and meter every thought, or do you own the agent and rent only the record?
"Breeze is not a fix for a poorly maintained CRM," one honest critic wrote. "It is an accelerator for a well-maintained one. It accelerates the mess." I believe that. Which is exactly why the right answer isn't "adopt the wall" or "ignore the hype." It's to walk through the door — own the agent, keep the mess under your own control, and pay nothing per action.
The wall is for teams that don't know the difference. The door is for the ones that do.
The question
So when someone at your company says "should we turn on the AI agents?", the answer isn't yes or no.
It's: inside or outside? Are you renting the agent's thinking and metering every action, or owning the agent and renting only the record?
HubSpot just shipped both answers in the same release and named the whole thing "Unbound." The decision — and the invoice — is yours.
I know which side I've been living on. The only question left is whether my day job will catch up to my night life.