A weekday, end to end
Most of my work happens before I open the laptop.
First drafts of research, analysis, and specs arrive in hours, from agents I built against six live data sources, on a schedule that runs whether or not I’m awake.
An ingest agent reads the day
Every note I took, cross-linked against everything before it.
Scheduled agents have reported
Six live systems pulled into one brief, every claim cited.
Agents take the first pass
Research, analysis, specs, prototypes. That is how the first version gets made.
A curator agent mines the day
It proposes changes to the system itself. I approve them in the morning.
I was the only PM at a 50-person lender, so my own capacity capped what product could do. That is what pushed this from a habit into a system I maintain, and then raised the obvious next question: what happens when everyone else is running agents too?
Everyone wanted more automation. One colleague alone was running forty agents.
Every new agent made the company more fragile, because no agent could see a whole loan. Trigger-based automation scales with the rules you maintain. Context-based agents scale with your data model. So I paused new automations and rebuilt the layer underneath first.
What I built for the company
A context layer every agent could read
One per deal, assembled on demand, so context survived every handoff. Run against closed loans, it caught an accrual error on a funded deal.
Systems unified5 → 1 data layer
A rule for who may ship an agent
Split by permission, not by person. Read-only agents ship free; anything that writes to a system of record passes an engineering gate.
Gate applies tosystem-of-record writes
A standard someone else could build on
A written standard for how agents record changes, so two of them could not overwrite each other, and cheap enough to run that cost stopped being an argument against it.
First agent by a non-engineershipped in one sprint
What came of it
The point of building the layer underneath rather than more automations was that other people could then build on it without me. That is what happened, and it is the part I would point at. Servicing loan ops shipped their own agents for borrower outreach and for issue resolution on loan maturity and delinquency, running on the standard and the context layer, and cut the non-performing loan rate by about 10%. I did not build those agents. That is rather the point of a platform.
I should be straight about where it stopped. When my role was eliminated the first loan-processing agent was running in shadow operations alongside the team, not automating anything on its own. The foundation shipped and was in use; the automation it was built for was still ahead.
The same idea, pointed at my own job
Before any of the company work, I built the same kind of system around the product cycle itself: a context store that holds what the product and the company already know and drafts the PRD from it, prototypes built straight off that PRD, and a QA agent that checks shipped code and the interface against the spec.
Features through it in 2026
Specs, prototypes and QA all ran through the same context, so the second feature knew what the first one decided.
Features50+
Prototype turnaround
Stakeholders reacted to something they could click before engineering was committed, rather than after.
Fromweeks → hours
Design QA
The agent caught small bugs on complex work that two people reviewing by hand were missing.
Per week6 hrs → 1 hr