Revision · Project 5 — Forge

Forge crossed the boundary from workflow to agent. That did not remove structure; it made structure more important.

Carry these ideas forward

• Serious AI CLIs such as Codex are systems around a model, and Rust is used in a maintained real-world agent.
• Tool descriptions help selection; typed arguments, policy, and capability scope enforce behavior.
• Tool output is untrusted evidence even when it contains instruction-shaped text.
• Progress requires observable state change and new evidence.
• Terminal states and budgets belong to one loop owner.
• External writes need exact authorization, idempotency, reconciliation, and receipts.
• Causal traces enable diagnosis; strict replay performs no new inference.
• Durable memory is a verified write with scope, conflict policy, consent, and expiry.
• The deployable unit is a harness release, not one prompt string.

Defend the design

Choose one Forge tool with a side effect. Trace model proposal → argument decode → policy decision → approval → dispatch → receipt → state verification → completion evidence → replay. Any implicit arrow is a candidate bug.

Next project: Model Lab — understand, run, route, and improve models →.