ZERO Makes Agent Loop Explicit

The core idea makes sense, and the visible artifact trail is the strongest part of the project because it gives users something to inspect, rerun, and trust.

• –The planning/execution split is sensible if planning ends at a concrete, testable contract: files to create, commands to run, and success criteria to verify.
• –Execution should begin the moment the system starts mutating the workspace or invoking tools; anything before that should stay in planning or synthesis.
• –Reliability should come first: deterministic task state, retries, sandboxing, idempotent steps, and strong verification will matter more than adding broader capabilities early.
• –The project is already positioned as an engineering runtime rather than a chat agent, which is the right framing if it keeps producing auditable artifacts.
• –The main risk is scope creep: once planning becomes too free-form, the loop turns back into a generic agent and loses the value of explicit phases.