NO HYPE AI · THE DISPATCH
A Project remembers. An agent acts
From a Project that waits to an agent that acts, one real example.
A Project remembers. An agent acts
From a Project that waits to an agent that acts, one real example.
Paul Cobb
September 07, 2026
In a previous issue I showed you the Rental Manager Project I run. It reads my monthly statements, trends the maintenance, flags the CapEx, and hands a one-page summary to my CFO Project. Useful. But it still waits for me. I have to open it, drop in the statement, and ask. This issue is about the next rung: turning that same Project into an agent that goes and does the work, and the guardrails that make me comfortable letting it.
What actually makes something an agent. Opinion AI (emergingai.substack.com) put the cleanest definition on this I've seen, and it's worth crediting: an agent is not a chatbot with a longer prompt, it's a model placed inside a small working system. Their formula:
agent = model + goal + context + tools + state + loop + checks + permissions
The model is the intelligence. Everything around it is what turns that intelligence into finished work. Here's the part most people miss: a Project already gives you three of those pieces. The files are context, the memory is state, and the goal is sitting in the instructions you wrote when you set the thing up. That instructions file is what you're about to upgrade. Turning a Project into an agent is not starting something new. It's adding four things to a file you already have: tools to act, a loop to try and fix, checks to judge the result, and permissions to bound what it can touch.
The same Rental Manager, the same instructions file, upgraded piece by piece.
Goal: the finish line it already had, a clean monthly Ops Summary I can act on.
Tools: instead of me pasting in the statement, it reads the new statement straight from Google Drive where it already lands, and it drafts the PM email in Gmail for me to review. Those two connectors, Drive and Gmail, are the difference between a chatbot and something that acts on your real data.
State: the running log it already keeps, month over month.
Loop: this is the shift. It doesn't produce the summary and stop. It produces it, runs its own check, and if the check fails, it fixes and re-runs. It stops only when the work passes.
Checks: the check is the whole ballgame, and it's a line in the instructions, not a setting buried in a menu. Mine are specific. Every number has to tie to a line on the statement or get tagged UNVERIFIED, and it recomputes the PM fee against collected rent and flags any month that doesn't reconcile. The model proposes; the check is what I trust.
Permissions: this is where I keep a hand on the wheel, and it's three more lines in the same file. It drafts the PM email, it does not send. It flags any spend over my approval threshold, it does not approve. It appends to the log, it never rewrites history. An agent that can act can act wrong, so I bound what "act" is allowed to mean, in writing.
The play: add one tool and one check. You don't have to build the whole thing to get most of the benefit. Open the instructions on a Project you already run and add exactly two things:
One tool: point it at where the input already lives (a Drive folder, a label in your inbox) so you stop being the copy-paste step.
One check: write the test the output has to pass before it is allowed to finish. "Every figure ties to a source line, or it gets flagged" beats "make it good" every time.
Then keep the permissions tight: draft, do not send; flag, do not decide. That is an agent you can trust. It does the fetching and the first pass; you keep the judgment and the send.
Where it bites. Two honest limits. First, the loop is only as good as the check. A vague check gives you a confident agent that is confidently wrong. A specific check is what makes the loop worth trusting, so spend your effort there, not on a clever prompt. Second, resist the urge to hand it the keys. The point of permissions is not distrust, it is that the expensive moves, a sent email, an approved spend, should stay behind a human for a good while. Mine reads, drafts, and checks. The send and the spend are still me. That is not the half-built version. That is the version I want.
If you haven't built the Project underneath this yet, start there. That was last issue, and the instructions template is linked in it. Everything above gets written into that one file, which is the point. The Project is the thing that remembers. The agent is what that same file becomes once you give it hands and a leash.
Know an operator still copy-pasting into a chatbot every month? Forward it. That's how this list grows. And hit reply: what's the one check you'd make your agent pass before it's allowed to finish? I read every one.
Written by Paul Cobb, real estate operator (Louisville / Kentuckiana), forVEX builder. I run the plays I write about. More depth: forvex.app/journal