Hubsessed Office Hours - AI-Native - Oct 26, 2026

Nico LafakisHost
Trisha MerriamHost

Monday, October 26, 2026

Nico Lafakis, Trisha Merriam and Chris Carolan compared notes on working with AI agents: why Claude Code remembers what Claude projects forget, onboarding an agent by voice, one Dot or a team of agents, a riddle models keep getting wrong, and a checker agent that lifted accuracy.

Clips from this session

9 short cuts from this conversation.

Portrait

9 clips

1:22Claude projects forget; Claude Code remembers

1:05Breeze already knows not to make duplicates

1:20Migrating custom objects without thinking is a trap

1:22You don't hand a new coworker a dossier

1:21Talk to your agent the way you'd talk to a person

2:46One Dot or a team of agents?

2:20The surgeon riddle: a training-data fallacy

2:33The day we added Attest

1:24A HubSpot extension that thinks like a reporting admin

Key takeaways

Work kept in Claude projects tends to be forgotten; Trisha found Claude Code remembered far more, even for work that isn't code.

Before writing a rule into an agent's instructions, check whether the tool already does it: Chris reasons that Breeze, as a CRM administrative assistant, already avoids duplicate records on its own.

Hubsessed Office Hours - AI-Native - Oct 26, 2026

Clips from this session

Portrait

Key takeaways

Work kept in Claude projects tends to be forgotten; Trisha found Claude Code remembered far more, even for work that isn't code.

Before writing a rule into an agent's instructions, check whether the tool already does it: Chris reasons that Breeze, as a CRM administrative assistant, already avoids duplicate records on its own.

A migration that promises no remapping still needs someone to ask whether the process should work differently in the new system.

Claude Code looks to the folder structure first, so a simple, consistent set of folders gives every idea and every agent a place to land.

An agent that delegates has to pass on the context, or every agent ends up waiting on the human.

Onboard an agent the way a new coworker gets introduced: a short introduction and conversation, not a dossier on everyone.

Voice makes it effortless to change direction and share preferences as they come to mind, which gives an agent the reasons behind them.

One agent per person or a team of specialized agents: the choice shapes cost, context and how much the human orchestrates.

Some hallucinations may come from training data: a riddle answered wrong often enough gets answered wrong by the model.

Asking a model to hurry trades away checks and compute, and some words in a prompt raise its reasoning effort fast.

Chris found that adding a checker agent took the accuracy of call outputs toward 100%.

Memory needs judgment: the most recent decision is not always the one that should weigh most.