Loop Training Mode

A graduated-autonomy guardrail for a newly built Agent Loop: the first few times you run a new loop, force it to pause at every step and wait for human approval ("quick check before I burn the tokens") before it proceeds. Only once you have watched it and trust it does what you intend do you turn training mode off and let it run unattended. It serves two purposes at once — confirming the loop is actually doing the right thing (not silently drifting), and avoiding wasted tokens on a loop that is off-course — and a third, softer one: watching the supervised runs gives the operator a real understanding of the process rather than a black box. The move is progressive trust: autonomy is earned per-loop, not granted by default.

The complementary rule is for goals that aren't cleanly measurable: rather than pause at every step forever, keep human verification checkpoints at the few high-leverage decisions where a wrong turn ruins the whole run. The framing is "AI is like an intern" — you gate the choices that shape everything downstream (a party's date, venue, theme), then let it execute the rest. This is Agent Supervision applied to loops: relocate oversight to the run boundary and the pivotal decisions, not to every token.

Claims

• Run a new loop in training mode first — pause for human approval at every step until you trust it, then disable it. best practice — context: the first few runs of an unproven loop; trust is earned per-loop before autonomy is granted, and the guardrail is cheap relative to a loop that burns hours going the wrong way.

• Training mode earns its cost twice over: it confirms the loop does what you intend and it stops a wrong loop before it wastes tokens (and it teaches you the process). observation — the concrete payoff; corroborates Agentic Simplicity's "don't pay for autonomy that doesn't buy you anything."

• For less-quantifiable goals, place human verification checkpoints at the high-leverage decision points where a wrong turn ruins the run — don't try to verify every step, and don't verify none. best practice — context: non-measurable loops; the "AI as intern" rule — the more pivotal-but-unverified checkpoints you skip, the further the agent can drift. See Agent Supervision.

• Autonomy is graduated, not binary — a loop moves from supervised to unattended once its behaviour is verified, not on the day it's written. principle — durable: unbounded autonomy from step one is the failure mode; the same bounded-loop discipline as Agent Loop, applied to when you let go.

• Autonomy can also be claimed rather than earned, by making every update pass a strict held-out gate instead of a human. observation — the automated counterpart to graduated trust: SkillOpt runs unattended from round one because a rejected edit cannot land (see Validation-Gated Update). Training mode earns trust by observation; a gate asserts it by construction — and only over what the score covers.

Related

• Validation-Gated Update — the automated route to unattended operation; training mode is the human route to the same permission.
• Agent Loop — the artifact you run under training mode; the bounded-run/hand-back discipline is the same, here governing when supervision relaxes.
• Agent Supervision — the general practice: keep a human at the boundaries and route attention to the decisions that matter; training mode is its loop-specific form.
• Skill-Driven Loop Development — the build-time complement: pre-validate the skills; training mode is the run-time check that the assembled loop behaves.
• Agentic Simplicity — "don't burn tokens on autonomy you haven't earned" is the same cost-discipline seen from the loop-operations side.
• Self-Improving System — the "only trains chest" drift is why a self-improving system stays augmented (human signs off on direction) rather than fully autonomous.
• Distillate: Loop Engineering, Illustrated: Triggers, Skills, Verification, Memory.
• Distillate: How to Build a Self-Improving System with Claude Code — augmented-over-autonomous argued via the workout analogy; full autonomy removes judgment and drifts.
• Distillate: microsoft/SkillOpt — training skills like weights, without touching weights — the gate-not-supervision alternative for granting autonomy.

Linked from

• Agent Loop
• Agent Supervision
• How to Build a Self-Improving System with Claude Code
• This Week
• Loop Engineering
• microsoft/SkillOpt — training skills like weights, without touching weights
• Self-Improving System
• Skill-Driven Loop Development
• Loop Engineering, Illustrated: Triggers, Skills, Verification, Memory
• Validation-Gated Update