My background does not start on the technical side. It starts in business, marketing and sales, which is where I learned what an hour of the wrong work costs. The technical part came after, and it came because I needed it: automating something badly is worse than not automating it, so I had to understand what I was building to make it effective and safe.
What I built, over months, is a set of agents that run the parts of my week that repeat. This piece is the map of them. The full versions, with the lid off, live on the agents page; here I want to explain the idea, because the idea is the part that transfers.
The word moves
Agent is a word that changes meaning depending on who is using it, and the definitions are still being argued. Anthropic draws the line at systems where the model directs its own process and its own tool use, instead of running a path somebody wired in advance.
Mine is narrower, and it comes out of use. An agent is something that takes a piece of my week without me in the room, leaves its output where the next one is going to look for it, and gets better because every correction I make goes into the file it reads before it starts. If it does not do those three things, it is a script, and I have plenty of those too.
Six of them
These are the ones I run most weeks, named by what they take off my hands.
The one that finds what we have in common. Prospecting: it reads a company, a role and a person, and comes back with the reason to write and the first message.
The one that analyses your competitors. A gap and SWOT read with sources I can click, not a paragraph of adjectives.
The one that learns a brand by looking at it. Give it a site and a few files and it writes the brand system back to me, colours, voice and all.
One page, then every format. The deck, the PDF and the landing all come out of one HTML page, so there is one place to fix a mistake.
The one that does the research. Account intelligence before a call, with every number carrying the route to check it.
A CRM built out of spreadsheets. It reconciles what I did across channels, and it tells me what is due before I ask.
None of them talk to each other. They talk to the same folder.
One folder, not six tools
That sentence took the longest to earn, and it is the whole difference between owning six tools and running one system. Each agent writes what it found where the next one is going to look for it. The prospecting run leaves a file the research run opens. The research run leaves a file the message run opens. The CRM reads all of them.
And every correction I make goes into the file they all open before they start. Not into a prompt I have to remember, into the folder. So the tenth run is better than the first for a reason I can point at, and the reason is a line in a file with a date on it.
The tools around it are ordinary: Claude Code and Codex to run them, Apollo, Apify, Sales Navigator and Talent Insights to look things up, NetSuite and Looker Studio for the numbers, Boomi and MuleSoft because that is what I sell at Digitech. The system is not any of those. The system is the folder.
The week nobody pays for
The clearest thing I can say about what they are is what they do on the days nobody pays for. The same setup runs my own week.
5:30 am, the brief. Calendar, inbox and everything that changed on disk overnight, waiting on one page when I sit down.
8:30 am, the inbox, already answered. It works out where each thread stands and leaves the reply written. I approve it or change a line. It never sends.
12:00 pm, scheduling. I say who and roughly when. It finds the slot and writes the invitation.
5:30 pm, the close. It reconstructs where the hours went from timestamps and commits, then asks only about the gaps it cannot account for.
This page was built the same way: a small local server, a script that rides on top of the page so I can leave a note on the exact element I want changed, and an auditor that checks every note actually made it into the file. That last part has caught me more than once.
Four rules that survived
San Francisco makes one thing easy: the models change every few weeks and somebody at the next table has already run the new one. These four survived that.
Pattern recognition is where the model earns its keep. It finds the shape in two hundred profiles faster than I can. Choosing which shape matters stays with me.
The loop beats the prompt. Chasing the perfect wording stopped paying. The run fails, the failure is cheap to catch, and the fix goes into the file it opens next time.
A claim with no file behind it does not ship. They hallucinate hardest on what sounds routine. Every number carries the route to check it, and Python does the arithmetic, not the model.
Autonomy is priced by the cost of being wrong. Reading, drafting and filing run unattended. Anything that leaves my hands is written by the agent and sent by me.
What comes back
Time, mostly. Hours for photography, for the people I want to see, for the projects I take on because I want to be in them. And inside the work at Digitech the question turns into the same one with more riding on it: how much more can we get through, how much faster, without the work getting worse.
The list of what broke is longer than the list of what worked, and that list is the reason any of this runs at all. It keeps growing because the work keeps handing me problems I have not had before, and each one ends up as another of these.