Hubsessed Office Hours Recap: AI-Native, October 5, 2026

In the first Hubsessed Office Hours for AI-Native, Nico Lafakis and Trisha Merriam spend most of the hour on one attendee question. Here is what they taught that a HubSpot operator can use on Monday morning.

What to do with this. Before you rewrite a single instruction, decide whether the agent has one job or two. If it has two, split it and let the second agent check the first. If it has to stay one agent, put the work into its instructions and keep a growing list of rules. Where a wrong value costs you, add a checker. The rest of this recap is the how, in the hosts' own words.

Watch the full replay on the session page. Times below are the replay's own clock.

The question

An attendee asked: how do you get a HubSpot agent to follow a template the same way every time?

Trisha had raised the same worry from her own work just before the card. Give an AI a list of rules, she said, and it ignores a small handful of them, a different handful each run. Nico's answer took the next forty minutes, and it came in layers.

Monday morning, step by step

1. Decide whether it is one job or two

Nico's first test is a toolbox. "If you think about agents as tools in a toolbox, that helps you understand the answer to this question." (Nico, 14:51) If you would ever use the two functions apart, or do not want them living together, give each its own agent and run one right after the other.

His own answer for how many: two. The first agent does the synthesis and outputs the values you care about. The second picks those values up, checks that they line up, asks the first agent again if they do not, and only then writes them into the fields.

His reason is about how an agent loses its place. By the time it has taken in a long run of material, the instructions it started with are far behind it, and Nico's view is that once they have dropped off, they have dropped off. His rule of thumb: "…one job per agent and that should definitely solve the issue." (Nico, 21:20)

2. If it has to be one agent, work the instructions

If everything has to live in one agent, Nico says to spend what he called an obscene amount of time on its instructions. He opened a customer lifetime value agent he built and walked through how it is put together:

• Markdown headings signal importance. H1 carries the most weight, then H2, then H3, with paragraphs for the detail. He added that some models follow markdown and some follow JSON, and that the formatting can change how strictly instructions are followed.

• Role first, goal second. "It's written like this on purpose, in terms of what the role is first, what the goal is second." (Nico, 26:54) Trisha pointed out that the goal setting at the top of an agent in HubSpot is a different thing, tied to HubSpot itself, such as updating a property, and not the purpose of the job.

• Scope, execution mode, method. Nico described scope as the sandbox, where the agent is allowed to play. Execution mode is how it plays, and method is what it does while playing.

• Write it the way you would brief a person. "I always say talk to like they're a person. And it helps to remove essentially ambiguity." (Nico, 29:58) Define what each value means and how the values relate, so nothing is left to guess.

3. Keep a rule list, and grow it from testing

Nico keeps what he calls a rule set, or a constitution: the guardrails, the things the agent should not do. Where it comes from matters more than what it contains. "…every time something pops up that you don't want to have happen, add it as a rule. Start testing again, make sure that that gets filtered out." (Nico, 26:20) Sometimes the rule works. Sometimes it needs different wording, and he said wording matters quite a bit.

4. Be careful with examples, and describe the template when you cannot give many

Examples are Nico's best route to a fixed output, with a catch. "…if you can't give a good multitude of them, it will treat one or two examples as the only one or two ways in which you want to see that data." (Nico, 30:46) If you cannot supply several, describe the template in words as precisely as you can: for a table, what goes in the columns and the rows, and for a report, what becomes a header and where each section sits.

5. Add a checker where a wrong value costs you

Trisha's own setup for content uses two agents. "So I have, when I'm creating content, two different agents. One is the writer, and then one is the checker." (Trisha, 22:43) The checker tests the writing against every rule and sends it back until all of them are met. Nico agreed that some kind of supervisor, finalizer or reviewer is how it should be set up now.

The hosts did not fully agree on whether you always need one. Trisha asked whether testing and refining one agent's rules could replace the second agent. Nico said yes for his customer lifetime value agent, which returns single values and not a written report, and said that before instructions could be customized he would have split it into several agents with one checking the rest. Where the work is client work, his advice was plain: "…if you're building this for a client, … build security guards." (Nico, 34:05)

He suggested running the reviewer as a workflow action, which he said he believed was required, so one reviewer can check several agents. Two cautions from him: test what each run costs in credits, and watch what happens to the cost when several agents feed the same reviewer.

6. Read a skipped rule as forgetting before you read it as disobeying

Trisha asked whether very long instructions make an agent break its rules. Nico would not call it that: "…if you forgot to do a rule, I wouldn't say, Trisha, you were disobeying the rules. I would say you forgot one." (Nico, 36:20) The output is still not what you expected, but the fix points somewhere different: how much the agent is carrying, and how the job is split.

Nico also said he expects that after December, splitting agents will not be needed. That is his expectation, offered as such, and not something either host tested on the call. The hosts' summary for today: have a reviewer, build a rule set, think about guardrails, and test, test, test.

Also on the tape

Agents talking to agents. On connecting your tools, Nico drew a line he cares about: "…with an MCP connector, you're allowing your agent to open the door and go in and take a look at everything. With an agent-to-agent connector, you're allowing your connection only to go talk to the other curator, let's say, of that information." (Nico, 10:48) Trisha, who supports more than eleven HubSpot portals, asked whether one assistant could hold all her tools without disconnecting and reconnecting, and the hosts talked through Dots as one possible answer.

Which model runs under HubSpot's agents. Trisha suggested asking Breeze, and the hosts tried it live. The tape does not settle the answer. Nico would like to choose the model, as some other apps allow, since that would also set the credit cost.

Reverse engineer your data. Nico's favorite exercise for anyone who loves data: "Whatever it is, reverse engineer it. I don't care what it is. Whatever the data set is, reverse engineer it." (Nico, 47:50) His example was how deals are shared among sales reps. Start at the end: a deal closed because of a relationship, so look at each rep's relationships and won deals, and find the threads the customers share. Trisha has done this by hand with call transcripts and said transcripts inside HubSpot make it much easier now. She has seen it show that one person was the best fit for one product, and in other cases it became a coaching opportunity. "I love the reverse engineering. It's so illuminating." (Trisha, 54:02) Nico's line on the handoff between reps: "…deal split is better than deal lost." (Nico, 53:18)

Next session

Nico and Trisha are back next Monday, October 12, in the same lane. Bring a question about HubSpot AI, agents or automation: Hubsessed Office Hours, AI-Native.