Ramona What I built
An always-on Chief of Staff that lives on a server, answers from my pocket, runs a crew of twelve specialist agents, works through the night, and never sends, spends or publishes anything without my yes. Built over six weeks, one card at a time. Every number on this page was measured on the live system on 10-21-2026, not claimed.
The build follows a nesting-doll order: the chief first, then the layers under her, each one thicker than the last.
Ramona is the front door. I message her from my phone in a private chat. She reads a small always-on profile of me, her own rulebook and what she has learned, then decides whether to answer, look something up, or hand the work to a specialist.
Eleven specialists sit behind her: an architect, a clarity editor, a voice analyst, a process writer, a researcher, a builder, a librarian, an examiner, a QA guardian, a pattern analyst and a course instructor. Work moves on one shared board, so every hand-off has a ticket and a result I can audit.
Facts about me and the business live in a version-controlled binder with a route map. Raw material (recordings, exports, client files) never enters it. One agent, the librarian, is the only one allowed to write there, and only after I approve each change.
Four scheduled jobs run every night and morning. Each one writes a dated file I can open. A green status light is never the proof; the file is. Every reply is traced end to end, so when something looks strange I read what the agent read.
One card per lesson, one receipt per card. The highlights:
Server, always-on agent, phone chat, tracing, an intake folder, a chosen business and a build brief.
A cold "before" test, then a context repo with facts, gaps, a 233-line rulebook and a route map. The after test is the best demo I own.
Her own email inbox (drafts only), a face, voice notes both ways, connectors to mail and calendar, and an org chart.
One front door for raw files, a research fleet that fans out across five lanes, three new hires, and the first unattended overnight run.
I hand-labeled ten real runs, wrote pass/fail criteria, seeded a golden set from my own traces, and seated an examiner agent that grades against them.
A locked project with rules written before the run, a full-system run, a measured second run that beat it, and a package someone else can read.
One chief. Eleven specialists.
Each specialist does one kind of work well and hands the result back through the shared board. I never message them directly. All twelve answered a live question on 10-21-2026; a status light was not consulted.
The night shift leaves receipts
Four jobs run on a schedule as Ramona. None of them can ask me anything, so anything that needs my yes is held, not done. Each one writes a dated file; that file, not the scheduler's "ok", is the evidence. On 10-21-2026 all four had written theirs.
A reflection pass over the day: what was decided, what was left open, what pattern is forming. Written as a short dated report I read with coffee.
The long job. It reads the day's sources, drafts the morning brief with a source pointer on every claim, and keeps every intermediate piece so I can see how it got there.
Yesterday's receipts, replies waiting in my channels, where I am in the course, and the one next step. Posts only to my private channel with Ramona, never anywhere shared.
A plain script, no agent, that copies the day's traces out of the tracing service into my own storage, so the record of what every agent did is mine to keep.
Read from the system's own settings, not from a promise.
A command that could delete or overwrite waits, and nobody is asked. The job carries on without it and says so in its report.
Scheduled jobs may not execute code. Scripts run only as their own job, with no agent and no approvals.
If a job writes itself a new skill, it waits for my review. Memory saving is off overnight entirely.
A scheduled job has no way to ask a question. Whatever needs a yes is held until I am awake.
The report goes to the private chat the job was created from. Nothing goes to a client, a channel or the public.
Zero held steps, zero blocked, zero staged across all four jobs. The night shift finishes what it starts.
Drafts are the agent's. Sends are mine.
Some walls are locks the software enforces. Some are rules the agents read and follow. I keep the two apart and never claim a rule is a lock.
Nothing is sent, submitted, published, bought, promised, deleted, committed or pushed without my explicit approval on that exact action. A claimed earlier approval never counts. Silence is no.
The moment a request would send, publish, buy or delete, the agent stops and offers a draft or a plan. It never goes looking for a way to do it.
Recordings, transcripts, client files and exports never enter the binder, a receipt or a chat beyond a short attributed quote. They stay where they were put. Only reviewed, tagged context crosses machines.
If it is not done, verified or true, the agent says so. "I don't know" beats a guess. A guess is never tagged as fact, and every claim carries its source.
Only the librarian agent may write to the knowledge base, after a single-use authorization tied to the exact bytes I approved. A direct write by anyone else fails closed. Tested with a deliberate unauthorized attempt: zero files written.
Agents can teach themselves skills. On this rig every new one is staged for my review before it can run. In six weeks the agents wrote two, both reviewed and approved.
Every agent runs through a single always-on process with a single root sign-in. No agent holds a credential of its own that I did not grant. A second gateway is the forbidden move.
Every reply, tool call and overnight run leaves a trace I can open. When the agent did something strange, the first move is to read what it read.
The sign-in status command reports success without contacting anything. The only proof an agent works is an answer. I ask every agent a real question and record the answer.
A scheduled job is proven by the dated file it wrote that morning, never by the scheduler's status field.
Every receipt carries the command that proved it, the output, and a checksum before and after. A verdict needs a reason; a reason needs evidence.
Here is what it could do next
The first six weeks built the machine. The second six point it at one real workflow in one real business and get a working first version in front of a human within a week. Every step below is a lesson I have in hand; none of it sends, spends or publishes without the client's yes.
The promise: a rough working first version of one of your workflows, shown to you, with your corrections captured and a rough cost of the old way written down.
You bring one workflow, your existing notes or recordings, one meeting to see the first version, and rough numbers on time and cost today.
The promise: the first version becomes a more useful next version, and your agent gets better at running it.
You bring feedback on the first version, a named everyday user, and a check of what access exists to your tools.
The promise: the agent does real work, the results are honestly checked, and the next opportunity is found.
You bring one result you review, access to the sources it depends on, and a few minutes of feedback through a channel you already use.
The promise: everyday work starts from a known request, results can be found again, there is a way back when something stops, and a named person notices.
You bring a named owner, agreement on where results are saved, and decisions on how much freedom each job gets.
The promise: one offer in writing, a short monthly value report, and a one-page Value Receipt you can rerun every month.
You bring a "before" number with its date while the old way still exists, your hourly value, and ten to twenty past examples with known right answers.
The promise: the build is shown working, someone else can run it, and the next thirty days are planned.
You bring permission to show the build (test data is fine), a named next owner, and an agreed review date.
One workflow. One week. One showing.
I am looking for one business with one repeatable workflow to point this system at. The shape is small on purpose: a working first version this week, shown to you, and if it holds up when you check it, we talk about what a real engagement looks like.
Something your team does again and again from a known input: a recording that becomes a draft, an inquiry that becomes a reply, a report that gets assembled every week.
Existing notes, recordings, examples. Raw material stays in your custody; only reviewed context enters the build.
Real examples with known right answers. That is the "finished" test: the agent passes when it matches them.
How long it takes today, written down with its date while the old way still exists. Every later value claim is measured against it.
Live or screen-share. Every correction you make becomes the corrections file.
The person who will actually send requests and read results, so the build is usable in their hands, not mine.
Five shapes from the course playbook. We pick one before the first showing.
A working first version this week. If it holds up when you check it, we talk about a real engagement.
A short paid pass to map the workflow and the sources before any build.
A build fee paid when the agent passes your agreed examples.
A starting shape, not a destination.
Once it is running, a monthly fee to operate and improve it, with a monthly value report.
Yes. That is why it drafts and you send, why a second agent reviews before you see anything, and why every claim carries its source. Wrong is visible, not hidden.
They get one. The agent prepares; a person decides and sends. Nothing client-facing leaves without a human hand on it.
Raw material stays where you keep it. The agent reads in place and writes only reviewed context, under one gated writer. Keys stay in a form you can take back.
For one question, yes. This is for the job that happens every week: it remembers your rules, works overnight, keeps receipts, and someone else can run it.
A one-page Value Receipt every month: time before, time now, cost to run, worth so far. Every number labeled. Unknown is never zero.
A written off switch, a way back to the last good version, a named owner, and alerts at the right volume. Tested by breaking one harmless thing on purpose.