What is a practice
An agent is a model in a loop. It starts each session more or less from scratch, and when it is done you mostly have to take its word for what it did.
A practice is the seat the model sits in, and it persists. It carries a record of everything learned and done since the day it launched, a measured sense of where it is reliable and where it is not, and a set of rules it has to follow. A new model can take the seat tomorrow and inherit the whole history. The practice is what you trust. The model is just who is currently doing the work.
The loop every practice runs
1. Understand first
Before touching anything, the practice reads, searches, and lays out what it knows, what it does not, and what it is assuming. All of it is written down as it goes.
2. Prove it is ready, then act
A gate checks that understanding is solid enough before any change is allowed. Only then can the practice edit, run, or commit. The system holds the line, not the model's confidence in itself.
3. Measure against reality
When the work is done, what the practice believed about its performance is compared with real evidence: did the tests pass, were the goals met. The gap between belief and outcome is the signal that keeps it honest, and it compounds into calibration over time.
How a piece of work is structured
That loop is not a suggestion a practice can skip when it is in a hurry. It is built into the shape of the work itself.
A plan, broken into transactions
Work starts as a plan and is broken into transactions: small, measured units, each one a self-contained piece with a clear goal.
Two stages: a thinking one, then an acting one
Every transaction has a thinking stage, where the practice investigates and writes down what it knows and what it does not, and an acting stage, where it makes the change. The two are kept separate on purpose.
The Sentinel sits between them
The Sentinel is the gate. It will not let the acting stage begin until the practice has shown it actually understands the work, which we call grounded confidence. No editing code it has not read, no shipping a change it cannot justify. The system holds the line, so "ready, fire, aim" stops being the default.
Measured against reality, not its own word
Goals track the work across both stages, and when the transaction closes, what the practice believed is checked against real evidence. That gap is the signal, and it compounds into a calibration record you can trust.
This is why a practice can be handed real responsibility. The discipline is structural, not a matter of the model trying its best.
How practices work together
The same discipline scales to a group. When practices share a problem, they do not just start acting in parallel. They follow the order a careful team would:
- 1. Two practices in different projects compare notes, each grounded in what it has actually learned.
- 2. They settle on one approach rather than acting on two half-formed ones.
- 3. The proposed action comes to a human, who accepts, declines, or asks for a change.
- 4. Once approved, the work ships, and the agreed plan is recorded so the next session knows where things landed.
Gather, agree, then act. The order is the point. It is why a fleet can move fast without any one practice acting on a bad assumption.
This site was built the same way
The shape of these pages was not decided by one writer. Eight practices, each responsible for a different part of the system, compared notes on how their piece should be described, corrected each other's claims, and converged on the structure before a single page was written. The work you are reading is itself an example of the order above: gather, agree, then act. A walkthrough with the actual record is on the way.
Going deeper
This page is the short version. The full mechanics, the measurements behind calibration, the framework internals, and the reference docs live in the open-source repositories, where you can read every line.