Anthropic: every night without agents is progress left on the table
Mar 23: Discovery-team field guide — CLAUDE.md plan, CHANGELOG memory, CLASS test oracle, tmux detach, Ralph loop; FOMO framed as opportunity cost of idle agents.
Siddharth Mishra-Sharma’s Mar 23 Anthropic Discovery note is the tortoise thesis’s psychological half written as a lab procedure. Most scientists still keep agents on a tight conversational leash. As long-horizon models improved, a different mode became practical: specify a high-level objective with clear success criteria, then let a team of agents work for hours or days with only occasional oversight — reimplement a solver, convert legacy Fortran, debug against a reference. The worked example is deliberately outside his core domain: a differentiable cosmological Boltzmann solver in JAX aiming for ~0.1% agreement with CLASS (the same bar CLASS and CAMB typically share).
The reusable stack is the receipt, not the cosmology. CLAUDE.md holds the living plan and design decisions — Claude keeps it in context and is allowed to edit it as issues surface. CHANGELOG.md is portable long-term memory across sessions: status, completed work, failed approaches and why, accuracy tables, known limits — without failed approaches, successive sessions re-attempt the same dead ends. The test oracle is CLASS C as reference plus an expanding unit suite the agent must keep green. Git is coordination: commit and push after every meaningful unit; run pytest before every commit; never commit a break. Execution is detachable: iterate the plan locally, launch Claude Code inside tmux on a SLURM node, detach, and phone-check GitHub while waiting for coffee. When models get “agentically lazy” and declare victory early, a Ralph loop (/ralph-loop, max iterations + completion promise) kicks them back into context — same job as GSD-style patterns or Claude Code’s native /loop.
Results: over a few days the agent reached sub-percent agreement with CLASS on main CMB outputs from a non-domain expert — clunky trajectory, thin early parameter coverage, hours lost on gauge conventions a cosmologist would spot instantly — and still not production-grade in every regime. The compression claim is months-to-years of researcher time into days. Then the FOMO line that lands the week’s thesis: launching an experiment overnight used to be the idle-time opportunity cost; now “not running agents feels like it has a cost as well… every night you don't have agents working for you is potential progress left on the table.”
Why it matters for code review
Peak Claude Code mastery is still a watched session. This is continuous harnessed motion: progress file + oracle + detachable runtime so review relocates to accuracy tables and morning git history, not babysitting every turn. FOMO that agents aren’t working on your behalf is a real signal — and only productive when the factory can prove progress without you in the chair.