From Jeremi Karnell · November 2026
Meet Scout: The Agent Invited Last, on Purpose
HnyComb's first agent was the last one invited. First the data had to be reliable, then agents had to be able to talk to each other, then there had to be rules. Only then did this busy little bee get past the door.
Jeremi Karnell · 16 Nov 2026
Meet Scout. He's a scout bee, and his whole job is to fly out ahead, find what's worth finding in your numbers, and come back and tell you about it.
Quick confession before we go further. When this note first went up a few days ago, it was called Meet Ledger, and the agent was an owl in orange vanity glasses. Since then the brand has moved. ValuationMCP is now HnyComb, at hnycomb.ai, and an owl living in a honeycomb never made a lot of sense. So Ledger retired and Scout moved in. Same agent underneath, same rules, new wings. The Heath jokes retired with the owl, which is probably for the best.
Scout is also the first agent inside the app, and he was the last one invited to this party. That was on purpose.
He needed three things before he was allowed to touch anyone's practice: data he could reach, a way to talk to other agents that doesn't involve emailing spreadsheets, and a written list of things he may never do. Those came first, in that order. Only then did he get past the velvet rope to hang with the cool kids.
First, the data
In August, ValuationMCP (now HnyComb) became an MCP-native app. In plain English, a valuation runs inside Claude or ChatGPT, and your assistant reads from the systems that run your practice (the custodian, the CRM, the accounting software) via Model Context Protocol (MCP).
Then, two agents in one room
In October, the app became A2A compatible. A2A stands for agent to agent. It is an open protocol that lets one company's agent ask another company's agent for something, with a defined handshake. Google started it and donated it to the Linux Foundation, where Microsoft and AWS back it too. So an enterprise agent built on Google Cloud, Microsoft or AWS can reach the HnyComb request broker in a language it already speaks.
The first use is in M&A. A buyer's agent asks for what is missing. The target firm's agent answers from its own record. A person approves every step, and nothing is drafted or shared until there is an NDA. The back-and-forth that used to take weeks of email starts to look like a short queue of cards.
Then, the rules
The next piece was a harness. As far as I can tell, it is the first one defined for practice management, succession and M&A in wealth management. A harness is the set of rules around the model: what it may read, what it may suggest, what waits for your click, and what it may never do.
Every action an agent can take is registered in one of four lanes. Read runs freely. Propose becomes a card on your desk and nothing happens until you act. Approve runs only after your explicit click. Never is refused and written to the log. One rule sits under all four. The model writes the words and the engine owns the numbers.
I wrote the long version of that last week in A2A Needs a Harness.
Only then, the invitation
The agent showed up the day after the harness page went live. He didn't have to invent a single rule of his own. He inherited all of them, and he survived a full rebrand a few days later without losing one. Honestly, he's more cheerful about it than I'd be.
He's also optional. Scout is the agent that lives inside the app, for anyone who wants one already there. But HnyComb is MCP-native, so you can keep using whatever assistant you already use: Claude, ChatGPT or something else. Bring your own or borrow the bee. Either way, the engine owns the numbers.
He explains your number. Ask why your valuation is what it is and he walks you through it in plain English. Every figure he uses is inserted by the app from your saved valuation, never produced by the model. He puts it better than I do: "I'm a bee, not a pitch deck."
He does the legwork. Point him at a move and he works it, on a schedule if you like. He fills drafts from your records and flags the gaps instead of filling them with a guess. You review what he filled. Then he turns it into dated checklist steps and, when the work needs a professional, prepares the brief for your attorney, compliance team, accountant or insurance professional.
He writes the pieces. Client letters, buyer briefs, the documents a move needs. Every one is a draft for you and your compliance team to review.
He knows his limits. Reading changes nothing, so he just does it. Anything that changes your record, he asks first. And a few things he won't do at all, however nicely you ask. The Never column on his page includes "Bump my revenue a little so the number looks better." He declines. Politely, and every time.
On the Network plan he can also answer questions about network-level data, from saved outputs only and inside the consent each member gave. Under the fuzz, he runs on Claude Opus 5.5 for conversation and Claude Sonnet 5 for the loops.
Why he was invited last
Agent-first would have drawn more attention. Charming assistants demo well. Permission lanes never have. In this work you fail in two ways: a number stated with more confidence than the data supports, or a message sent that should not have been. Build the agent first and you have a likable way to do either.
Scout came last, so the data, the protocol and the rules were already in place when he started. He reads through the same MCP connections, follows the same four lanes, and writes to the same step log as every other loop in the app. That's also why swapping an owl for a bee took days and not months. The costume changed. The plumbing didn't.
Real scout bees work this way too. A scout finds a promising spot, flies home and dances about it, and the swarm doesn't move until enough of the colony agrees. Scout can find the gap in your draft. He can't turn that draft into a sent email.
What he does not do
Scout is not an appraiser, and nothing he produces is an appraisal, a fairness opinion, or investment, legal or tax advice. He does not replace your attorney, your CPA or a buyer's diligence. He never changes a figure on his own, never contacts a client and never sends anything without your say-so.
Last one in, life of the party
Scout was the last one invited, and he's the first one people will notice. Under him sit an MCP-native app that reads from the systems of record, A2A so agents at different firms can work a request with a person approving each step, and a harness that decides what any agent may read, suggest, wait on or never do.
The rules were written for more than one agent. The deal loops already run under the same harness, and whatever comes next starts where he did: on reliable data, inside the lanes, with a person signing off on anything that matters.
If you want to meet him, run a valuation and click Ask Scout. He's probably already out looking.
Market commentary from public sources, quoted as each publisher released it. Not investment, legal or tax advice, and not an appraisal of any firm.

