Status: superseded by ADR-0011 in session 02. Originally deferred, 2026-10-26: the author postponed open decision D2 in session 01. It must be decided before example 03 is scaffolded.

Context

Each piece has a deterministic core: a standard-library Python agent in the shared vacuum world (ADR-0002). CLAUDE.md §4 also sketches an optional LLM-powered “today’s twin” per piece, the same agent type in its present-day form:

  • 03 goal-based: a coding agent in a sandbox whose goal is a test predicate (“all tests pass”)
  • 04 utility-based: best-of-N candidates scored by a utility that includes cloud cost
  • 05 learning: an agent that rewrites its own instruction file from a critic’s feedback

For 01 and 02 the twin can stay in prose (a guardrail as a reflex agent; a memory file as internal state).

A “yes” brings a model provider and model, credentials (an API key stored with sbx secret, or a subscription sign-in inside the sandbox), network egress to the provider only, per-run model costs on top of sandbox minutes, and nondeterministic output. The articles would report one recorded run rather than promise identical results.

Options considered

  1. Yes, for some or all of pieces 03–05, with a provider, model and budget cap.
  2. No. The twins appear only in prose.
  3. Decide later. Pieces 01–02 don’t depend on the answer.

Decision

Option 3: decide before example 03 is scaffolded (/new-piece 03 checks this ADR).

Consequences

  • Pieces 01 and 02 are built with deterministic cores only. Their “Why it still matters” sections describe the twin in prose.
  • No agent_llm/ directories, no ANTHROPIC_API_KEY secret and no model egress for now.
  • When this is revisited, the answer must cover: which pieces, provider and model, a budget cap, and the authentication path.