sovereign-agent

Ruling: build the complete always-on teaching agent

Created: 2026-09-07 · Last-updated: 2026-09-07 · Status: ACTIVE

The book promises Build Your Always-On AI Agent From Scratch. The implementation owns the model/tool/observation loop and its runtime. Delegating an entire assignment to a CLI agent remains optional and is taught as a different abstraction.

Authorize one self-contained Telegram adapter, a bounded MCP client, versioned local skill files, an explicitly invoked service path using a standard OS supervisor, and one optional container-isolated teaching tool. These capabilities ship in this distribution. ZeoCore is optional and supplies maintained production integrations; no chapter requires its installation. This replaces the corresponding 1.x exclusions in durable non-goals. It does not authorize a marketplace, general workflow language, fleet scheduler, secrets product, or silent daemon installation.

The existing 7,250-line core ceiling cannot cover this authorized scope. Its successor must count all installed Python source, disclose the core subtotal, and retain an enforceable measured ceiling. Establish the revised allowance from implemented chapter checkpoints; do not conceal required adapters in an uncounted package or advertise the former 6,400-line figure.

Keep all thirteen existing chapter labs executable while adding the sixteen-chapter build sequence. Preserve source history and reader URLs through an explicit migration map. Publication requires runnable cumulative checkpoints, adversarial failure experiments, mechanical editorial gates, and a human sample reading. A score is an instrument, not a publisher endorsement.

Purchasing exercises use a separately stored simulated supplier. No live purchase is required. Unknown remote outcomes must remain unknown until discovered. Worker identity, human approval, and external-effect recovery remain separate contracts. Application policy is never described as an OS sandbox.

Verification: the project gate must exercise the reader-owned loop, durable intake and scheduling, exact-order approval, ambiguous supplier response, stale-worker rejection, isolated execution or honest refusal, and backup/restore. Scope and budget changes also appear in CHANGELOG. Independent review is requested in org issue 603; implementation continues while it is pending.