Riding the loop: a development mode where the operator rides the cycle rather than gating it

[aaronsb]

Operator-proposed methodology capture (2026-07-22), candidate for a future way. Not the only way to develop — one mode among several — but a mode that worked measurably well and has general shape beyond agent-ways.

The name

Human-in-the-loop casts the human as a gate: the machine's cycle stops, waits for approval, resumes. Riding the loop inverts the posture — the operator is a live user of the artifact while it evolves, steering by continuous lightweight signal from inside their own working surface. The loop never stops at them; they ride it. (Operator's wordplay, worth keeping.)

The cycle

try it out → notice something → utter it (one sentence, in the surface you're already using) → it lands with the builder as a recorded finding → implement behind a gate scaled to blast radius → deploy as a non-event → try it out again.

Preconditions (the general form)

1. A zero-friction feedback channel from the operator's usage context to the builder. Here: the attend bus + ADR-172 turn-boundary drain — an utterance typed mid-use reaches the building session without a context switch, ticket form, or meeting. Generally: any channel where noticing and reporting are the same act.
2. Findings recorded in the operator's own words. Each observation becomes a first-class work item quoting the operator's framing ('make sure the time constant isn't buried'), so implementation answers the operator, not a paraphrase. The record IS the spec.
3. A review gate that scales with blast radius. Structural changes get full adversarial review (today it caught two invisible-by-design defects: a cold-start message swallow and a destructive-intent inversion); display-only changes ride CI + self-review and land in minutes. A fixed-weight gate would either rubber-stamp the big or strangle the small.
4. Deploys cheap, fast, and reversible. Per-component versioning, hot-swap reload, projection reconcile — the artifact announces its own upgrade mid-conversation. When deploy is a non-event, 'try it out' resumes in seconds and the cycle's period stays short enough to hold the operator's attention.
5. The operator exercises the real artifact. Dogfooding surface = production surface; no staging copy whose lessons need translation.

When these hold, the bottleneck moves to the operator's attention — which is exactly where it should be, because attention is what notices.

Where it applies — and where it doesn't

Fits: artifacts the operator genuinely uses; decomposable changes; reversible deploys; a builder that can absorb interrupts mid-implementation (turn-boundary delivery made interrupts free rather than disruptive). Doesn't fit alone: one-way doors (they still want the stop-and-decide gate — the two modes composed today: ADR debates for direction, riding the loop for everything downstream of an accepted direction); greenfield with no live surface yet; work needing long uninterrupted design arcs.

Evidence (2026-07-22, agent-ways)

One session: ADR-172 and ADR-173 from live debate → Accepted → implemented → adversarially reviewed → released (attend 0.9→0.11, attend-chat 0.3→0.5) → deployed → operator-verified. Ten operator findings (#388–#394, #398, #400 + the tab-model spec accreted on #393) captured mid-use and landed the same day, several within minutes of utterance. The infrastructure that carries the loop (turn-boundary delivery, the channel surfaces) was itself built inside the loop it now carries.

Disposition

• Distill into a way (likely softwaredev/delivery or meta) once the shape has survived more than one day's use
• Cross-reference from the develop/merge/release skills as an alternative loop shape

[aaronsb]

Operator follow-up: this isn't one way — it's a candidate class of ways ('agent loops' or similar): a family whose members differ on where the operator sits relative to the loop and what the loop's period is. A sketch of the members visible from current practice:

Loop shape	Operator posture	Period	Exists today as
Riding the loop	live user inside the artifact, steering by utterance	minutes	this issue; ADR-172 delivery is its transport
Gated loop (human-in-the-loop)	approval gate at decision points; loop stops and waits	per decision	ADR debate → Accepted; one-way doors; /merge four-square human gate
Goal loop	sets the condition, then absent; evaluator judges evidence	until condition met	/goal + goal-author skill
Scheduled loop	absent; findings arrive as reports	cron period	/schedule routines, /loop
Orchestrated loop	commissions a fan-out, reads the synthesis	one shot, wide	Workflow/ultracode multi-agent runs

Each member way would answer the same four questions — where the human sits, what carries feedback, what gates changes, what the period is — so choosing a loop becomes a deliberate act with a vocabulary, not a default. Composition is the norm, not the exception: today ran a gated loop (ADRs) containing a riding loop (everything downstream), and the two never conflicted because the gate sat exactly at the one-way doors.

Naming candidate for the band: loops/ (softwaredev or meta domain), with this issue's content becoming loops/riding.