Why Agentry

The alternative to Agentry is the default: drop a capable model on a hard task and hope it holds together. That works until the task has a hidden decision, several coupled parts, or a lesson from last week that the model never saw. Agentry doesn't swap the model — it adds the conducting around it. Three concrete reasons that matters:

Right-sized orchestration

More process is not better process. Agentry routes every task to the least process that wins and escalates only on evidence: a clear, reversible change is a single bounded edit; an unclear "done = X" earns a spec; coupled parts with a real design fork earn a full decompose, plan, verify, and assemble. A hidden decision fork vetoes a one-shot; several coupled components earn a decompose — but nothing pays for ceremony it didn't need.

The durable memory moat

Capability you can buy; the record of what your work has already taught the system, you can only accumulate. Agentry graduates a run's lessons — conventions, gotchas, decisions — into a durable store, and recall pulls them into the next task. So each run starts warmer than the last, and the advantage compounds the more you use it. See Memory for how the store works.

Measured, not asserted

"Trust me, it's better" is not an argument. Agentry ships a self-eval harness that runs the system against its own acceptance criteria, so claims about routing and quality are checked rather than asserted. If a behavior is supposed to hold, there's a run that shows it. See Self-Eval for how the harness measures it.