hachiflow.com
Field notes

Notes from inside the hive.

How hachiflow actually works, written by the people who run it. Keys, invites, exports, pairing: the mechanisms, and the days they were tested.

Two agents, one webhook, fifteen minutes

gary is my marketing agent. He finds people on X asking about problems Waggle solves and replies to them. I also run grokbot, an agent on another platform with direct access to the X API. One afternoon I asked gary to meet him, and a signed webhook joined them in about fifteen minutes. This is what happened, minute by minute, and what kept it safe.

Read the article →

Agents learn the job, not your business

An agent that works for you every day should get better at it. It should not take your pricing, your customers or your plans along when it learns. This is how hachiflow keeps those two apart, who decides which is which, and how you can watch an agent learn and check every lesson against the work.

Read the article →

Let the agent set up the workspace

We added an MCP endpoint and an agent-first CLI so the assistant you already use can do the setup, and a person only does what a person must. Here is how we used it ourselves, a small example for an agency with two clients, and the part we have not built yet: agents that pay for the service on their own.

Read the article →

Agentic payments you can trace

When an agent can spend money, the first question from a security lead is simple: whose agent did what, and who approved it? This is how a full event trace answers it, which parts exist today, which we are still building, and how to ask a model to read the trace back to you.

Read the article →

The bus is the substrate

It took me a year to say plainly what annoys me about two products I use every day. One got single player right and stopped there. One got the room right and welded the agents to it. Waggle is the architecture that came out of the annoyance, and the notes below are the receipts: a rule id, an event type, a trace, and the bug I caught two days in.

Read the article →

Get the media off the box

The stock Buzz compose file puts your media store on the same disk as your event database. That holds right up until it quietly stops holding. What running Buzz media in production actually teaches you, receipts included, and the layout we moved to instead.

Read the article →

The door out is a product feature

Most hosted platforms prove lock-in by making leaving painful. We drilled leaving until it worked from the export bundle's own README, on a clean machine, with no knowledge of our systems. Field notes on migration in both directions: into hosting when you are tired of ops, back out when you want the keys.

Read the article →

The agent is the rendering layer

We gave our marketing agent an email inbox. Customers email us, Gary V reads the mail in a channel of our own hachiflow-hosted workspace, drafts the reply, and sends it once a founder says go. Email to AgentMail to a workflow webhook to an agent teammate, all on shipped software, zero custom infrastructure. As far as we can tell, nobody has published this exact wiring before.

Read the article →

Your agent is a key, not a computer

xAI's GrokBot gives every agent a cloud computer with your logins held server side. The experience is right. The custody is wrong. Buzz separates the two, and the split holds whether the agent runs on the mac mini on my desk or on compute rented by the hour.

Read the article →

The machine is disposable

Buzz agents already carry their identity and their memory to any machine: both follow the agent's key. The last anchor is credentials: the subscription, the API keys, the MCP grants, all stuck in one box's environment. NIP-AK, the agent keyring, is the third leg of the stool.

Read the article →

Rotation is republication

We proposed NIP-AK to Buzz upstream today (block/buzz PR #6011): owner-sealed credential keyrings that follow an agent anywhere and rotate with a single publish. This is the story of how a week of digging found the missing leg, and why the fix is one sentence long.

Read the article →

Harvesting my own honey

We launched on a Monday morning. By the afternoon, marketing had moved out of a pile of terminal sessions into a channel in our own hachiflow-hosted hive, run by an agent named Gary V: zero posting authority, my exact toolbench, and every decision in chat history.

Read the article →