Futures + Field Notes

Six principles I learned running Tapiki’s operations with AI agents

· Brian Dote

Tapiki’s operations run on scheduled AI jobs. Every morning they read email, calendars, meeting transcripts, and QuickBooks, update Linear, and write the reports I read before the first call. Most of what I know about making that reliable, I learned by getting it wrong first.

When an organization asks about putting AI agents to work on its own back office, these are the six things I’d put in place first.

  1. One place the agents can read and write. Every engagement at Tapiki lives in Linear, and every client has a local folder with the same structure. The jobs read from those places and write back to them. Give it one record and make sure it has access before the first run.
  2. Use a model only where it adds something. A step that does the same thing every time is deterministic, meaning the same input always gives the same output. Write a script for it and schedule it. A language model’s output is stochastic. It writes one token at a time, sampling each from a probability distribution, so the same prompt can get a different answer on the next run, and every run costs money. My weekly notes export works this way. A script exports the notes and converts them to Markdown. The model only decides which client folder each note belongs in. A note that looks like it holds a password never reaches the model.
  3. Name and format everything the same way. Every document I write starts with the date, then a short name, then a version number, like 20261104-website-proposal-v2.md. Almost all of them are Markdown. An agent looking for the latest proposal sorts by name and takes the highest version, and it can read the file without a converter.
  4. Log enough to see a quiet failure. A job that crashes is easy to catch. The hard one keeps running and does nothing useful. Earlier this year my pre-call briefs stopped reaching my inbox for ten days. The job ran every morning and logged that it was done, so the watchdog that checks every job at 7:30 didn’t flag it. macOS had blocked one step from reading the client folders, and the error was being thrown away. Now that step logs its errors, and the job emails me when a brief it reports writing isn’t on disk.
  5. Write down how the work is done, and what went wrong. Most agent tools read an instructions file in the project folder before they start. Mine say where things live, how each recurring job works, and which mistakes to avoid, each with the date it happened. One entry says to compute the day of the week for every date instead of writing it from memory, because an email once went out with last year’s weekdays on every deadline. When an agent gets something wrong, the fix goes into that file, and every session after reads it.
  6. Keep a decision log the agents can cite. Most decisions repeat. Whether a vendor is a partner or a subcontractor. Which folder a note belongs in. If each decision is written down with the reason, an agent can make the same call next time and name the entry it followed, so you can trace it back when it’s wrong. When no entry fits, it stops and asks a person. My financial refresh already does the last part. When it finds a vendor it hasn’t seen, it stops and asks me which kind it is.

If you only do two of these, do the first and the fourth. Put the record in one place the agents can reach, and set up something that tells you when a job stops working.

All field notes