We like to push how delivery works, building agents, testing new workflows, and turning what works into the way our Units operate every day. This is our workbench.
Thread #p-payments
Closed-loop control system: review, refine, re-run.
The system keeps looping until the senior moves it forward.
Model routed by difficulty
Dated, signed, and honest about the trade-offs.
Marcelo Núñez
Backend · domain
Aug 2026
Founding note
Model a system built with AI the way you'd model a thermostat: a closed control loop.
The agent produces the output. We read that output and correct it: tighten a skill, sharpen the context, adjust the routing. The value was never in typing the code. It's in judging whether what came out is true, and worth shipping.
Read the noteGustavo Iglesias · Damián Medrano
Delivery
Aug 2026
Experiment
We ran a public agent framework against agents we wrote ourselves. Here's what we kept, and why we're building our own.
There are good public frameworks, a whole crew of PM, architect, developer and QA, ready to run. We tried one. It's close to how we already think about a Unit. But renting someone's line means inheriting their opinions about how a story is written and how work gets dispatched.
Read the noteMarcelo Núñez
Backend · domain
Aug 2026
Lesson
A well-modeled domain lets AI write code that reads like a sentence. Changing that domain is where the bill shows up.
Give the AI a precise enough business rule and it writes code you can read out loud: a voucher is valid because it has at least one line, the concepts match, the totals match the lines. Secure on every side. Beautiful, and cheap to build.
Read the noteA growing library of proven capabilities shared across every Unit, giving our teams more leverage without sacrificing consistency, quality, or control.
A selection from the library · 18 of 45+
Reads a ticket, understands the backend domain, and breaks it into child tickets with standardized acceptance criteria.
Guardrail: nothing touches the database without explicit approval.
Rates each ticket's difficulty, routes it to the right model, and dispatches subagents to work features in parallel.
Routes: light → Haiku · bulk → Sonnet · hard → Opus.
One commit standard across every project: the story title first, then conventional commits, so any backlog reads the same way.
Reused: same skill, every repo.
Builds automated test flows through the Playwright MCP. Coverage and regression run as a standing function, not an afterthought.
Owner: a senior signs off on what the tests miss.
Bring a real backlog item. A senior runs it through the line and gets a first real commit into your repo inside a week. You watch the loop work on your own code, not a sample.