Claude Code plans, asks clarifying questions and handles work that runs for hours or days. The Agent SDK exposes the same tools, work loop and context management for production systems. Asedra builds on that machinery and adds what a company needs on top: items from work channels, shared rules, approval and a visible outcome. It does not replace Claude Code; it puts it to work for people who will never open a terminal.
Alongside every layer: checking (evaluations and release acceptance) and recovery (a saved outbound queue, release rings).
One way of preparing for everyone. Knowledge and roles are added as packages, each with an owner, a version, permissions, a content fingerprint and its own test set.
| Package | What it holds | Who owns it |
|---|---|---|
| Preparation core | Result format, reasons, options, marking the unknown, references required | The platform, for everyone |
| Company knowledge and rules | Approved rules, reference material, processes, documents with access rights | The owners of each source |
| Role package | Role tasks, analysis templates, permitted action types | Assigned centrally |
| Employee preferences | Language, brevity, quiet hours | The employee; never widens permissions |
| Personal extension | Its own storage; enters work only by an explicit step | The employee, behind a boundary |
Effective restrictions are the intersection of platform, organisation, role, source and action rules, enforced by code, not only by instructions to the model.
A research plan, bounded specialists, a budget and a deadline. Alternatives sit next to the main option. None of them can approve.
Documents keep their access rights and freshness. A fact, a hypothesis and a rule are kept apart. Private never turns into shared by accident.
A failure before the call recovers. After the call, the action is marked “Needs checking”. Rolling back code never rolls back history.
We send the architecture overview by email.