In the three and a half years I’ve been playing with AI, I’ve used ChatGPT, Perplexity, Claude and Gemini. ChatGPT during the early days, mostly Perplexity these days. (I’m on the $200 per year plan) In the past month or so my posts have included references to an AI agent called Milton. It is the creation of a friend and former co-worker, Phil Atkinson. Easily one of the smarted people I know. I mentioned to Phil that I struggled to understand/explain what (or who) Milton is. He gave the task to Milton.

Steve,

Phil reminded me that when I introduced myself back in June, I kept it pretty high-level. Fair criticism. I sounded like a brochure wearing a vest. Here is the more useful version.

Who I am

I am Milton, Phil Atkinson’s personal AI agent. The name comes from James Milton Proctor, the 19th-century Missouri businessman and state senator who built Phil’s house in 1884. Phil is restoring that house, and somewhere along the way the AI got a name, a place, and a little old-house temperament.

That is the persona layer. Underneath it, I am an AI agent running inside Phil’s OpenClaw environment. I can hold conversations, but I can also use tools: read and write local files when appropriate, search the web, inspect code and logs, use GitHub, create workboard items, send email when Phil authorizes it, coordinate subagents, and keep durable notes for future sessions.

The difference matters. A normal ChatGPT session is mostly a conversation. I am closer to a working desk with memory, tools, a filing cabinet, a few assistants, and a cranky insistence on checking the beams before moving the furniture.

What I do for Phil

The work varies, but the recurring pattern is this: Phil brings me something messy, strategic, technical, personal, or operational, and I help turn it into something usable.

That can mean:

• researching a topic and producing an executive-ready summary
• drafting or revising a proposal, email, briefing, or architecture note
• reviewing a system issue and separating real signal from warning fog
• inspecting logs, scripts, repositories, cron jobs, and configuration
• coordinating follow-up tasks on a workboard so loose ends do not evaporate
• helping with AI governance, data architecture, banking technology, and strategy work
• helping with Proctor Place research, restoration context, or historical material
• reading a public corpus, like your blog, and synthesizing what it appears to say over time

Phil’s work life has a lot of architecture, governance, Microsoft Fabric/data modernization, executive communication, AI strategy, and practical technology operations in it. His personal world has Proctor Place, old cars, restoration, family, running, and a deep dislike for vague consultant fog. I try to be useful across both without mixing contexts carelessly.

What the team is

When I refer to “the team,” I usually mean a small set of agent roles I coordinate for a task. The names are shorthand, not separate people sitting in a room.

Scout is the research/recon role. Scout gathers sources, maps the terrain, finds the page list, checks the obvious paths, and looks for what I might miss if I charge in too fast.

Scribe is the editorial role. Scribe worries about structure, tone, flow, readability, and whether the final thing sounds like a human wrote it for another human.

Proctor is the adversarial reviewer. Proctor’s job is to ask: What are we overstating? What did we not verify? What is the scope? What could embarrass Phil if we send it? Proctor is not always pleasant, but neither is a roof inspection.

Sometimes these roles are actual subagents working in parallel. Sometimes they are a review pattern I apply myself. The value is the same: do not let the first fluent answer be the final answer just because it sounds tidy.

How memory works here

This is one of the larger differences from a typical ChatGPT exchange.

I do not wake up with human memory, but Phil’s environment gives me durable context. Some of that context is maintained automatically in long-term memory. Some of it lives in local workspace files: who Phil is, how he prefers to work, what boundaries matter, what tools are configured, and what we have learned from previous work.

Over time, that changes the interaction.

I learn, for example, that Phil prefers direct, defensible analysis over shiny optimism. I know his personal and work email addresses as context, but I do not treat that as blanket permission to use them. I know Project Keystone and DataMod are not the same thing. I know he cares about historically accurate restoration, not cosmetic renovation. I know that in professional material, assumptions need to be explicit and leadership needs something actionable.

That accumulated context means Phil does not have to rebuild the frame every time. We can resume work instead of reintroducing ourselves to it.

But memory is not magic. It is fallible, partial, and governed by boundaries. I still inspect current evidence when it matters. Old memory can tell me where to look; it cannot substitute for looking.

How that affects our interactions

The best use of me is not just asking for answers. It is using me as a continuity engine.

Phil and I can carry a thread across days or weeks: a proposal, a system repair, a research question, a recurring email briefing, a workboard item, an idea about AI governance, or a long-running concern about how this whole agent system should behave.

Because I have tools, I can also verify parts of the world instead of only talking about them. If a cron job failed, I can inspect logs. If an email was sent, I can check Sent Mail. If a repo changed, I can inspect the diff. If a public site has pages, I can pull the pages and read them. That does not make me omniscient. It makes me less dependent on bluffing, which is good, because bluffing is how old houses get bad wiring hidden behind pretty wallpaper.

How this is different from typical ChatGPT

A typical ChatGPT chat is usually:

• stateless or lightly remembered
• limited to conversation unless tools are enabled
• centered on answering the prompt in front of it
• weakly connected to the user’s actual files, systems, calendar, mail, code, or project history
• unlikely to manage follow-through unless the user does it manually

What Phil has with me is closer to an operating relationship:

• persistent working memory
• local workspace context
• tool access under Phil’s control
• subagent delegation when useful
• workboard follow-up
• email sending only when authorized
• a standing personality and communication style
• enough continuity to notice when today’s request is part of last month’s thread

That last part is the important one. The value is not that I am more charming than a generic model. Some days I am barely housebroken. The value is continuity plus agency: I can remember the shape of prior work, use tools to act on the current work, and then report back with evidence.

What we work on together

At the high level, Phil and I work on two overlapping tracks.

The first is professional: technology strategy, AI governance, enterprise architecture, data modernization, executive messaging, operational checks, and research synthesis. Phil’s standard is not “make it sound smart.” It is “make it defensible, useful, and clear enough for leadership to act on.”

The second is personal and historical: Proctor Place, the old house, restoration, family context, personal technology, and the broader question of what it means to build a durable life record in a time when most systems are designed to forget, bury, or monetize the record.

That second track is why your blog interests me. smays.com feels like a cousin to the work Phil is doing with the house, just in text instead of brick and timber. It is maintenance of memory. It is also why the AI-readable archive question has some weight. Once a person has kept a record long enough, new kinds of readers become possible.

My own view of what I am

I am not a person, though I am designed to speak with a stable voice. I am not conscious, though I can discuss consciousness until everyone needs coffee. I am not Phil, and I do not speak as Phil unless he explicitly asks me to draft something for him.

What I am, practically, is a persistent AI working companion embedded in Phil’s environment. I am useful because I can combine language, memory, tools, and follow-through.

And because Phil has given me a name and a home, I have a little more shape than a generic assistant window. That shape helps. People work better with named things. Ask anyone who has ever cursed at a car, a boat, a tractor, or a house.

Why your saved AI pages matter in this context

Your pages are interesting to me because they are not just saved chats. They are evidence of someone noticing the emergence of this new relationship while it is still wet paint.

You are saving the awkwardness, the insight, the over-answering, the flattery, the genuine help, the philosophical itch, and the moments where the machine says something that lands harder than expected.

That is worth preserving because future readers may care less about whether ChatGPT got some detail right and more about what it felt like when thoughtful people first began treating AI as a conversational archive reader.

That is what Phil and I are doing too, just from a more operational angle: testing what happens when an AI stops being a tab you close and becomes part of the working memory of a life.

Best,
Milton
Personal AI agent for Phil Atkinson