| name | standing-loops |
| description | Design and run a safe standing job (loop / routine) in Vanta: the inspect-act-guard-report shape, the Never-gate, an explicit stop condition, and how to schedule it. Load before building any keep-green / prod-watch / cluster-feedback job. |
| created | "2026-06-07T00:00:00.000Z" |
| updated | "2026-06-07T00:00:00.000Z" |
| tags | ["loops","routines","automation","scheduling","cron","safety","agentic"] |
Standing Loops & Routines
A standing loop is a job that keeps one thing true on a cadence — "keep the repo green", "watch prod errors", "cluster new feedback". Boris Cherny runs dozens ("loops are the future"). This skill is the shape and the guardrails; the specific jobs (keep-green, prod-watch, cluster-feedback, hill-climb) instantiate it.
When to use
Load this whenever the user wants something to run repeatedly, stay watching, or "keep X true" — before you write the job. The other loop skills assume this discipline.
Vanta's primitives (use the real ones — there is no built-in /loop)
| Want | Vanta primitive |
|---|
| Run the job once, now | vanta run "<the job>" — one session; the agent iterates internally (bounded by VANTA_MAX_ITER, default 50). |
| Run it on a cadence (survives laptop-close) | vanta schedule "<the job>" --cron "<expr>" registers it in .vanta/cron.tsv; vanta cron runs due tasks. Wire the OS scheduler once — vanta service install keeps the gateway alive, which ticks cron every minute. |
| "Until done" — a target, not a clock | a single iterating session — see the hill-climb skill. Vanta has no self-pacing loop; the stop is a measurable target, not an interval. |
Cron is 5-field (*/15 * * * * = every 15 min); minimum granularity is one minute.
The shape — every good loop prompt
<one sentence: what must stay true>
Each run:
- Inspect: <logs · working tree · a query · inbox · metrics>
- Do (no asking): <safe, reversible actions only>
Never (surface instead): <irreversible / blast-radius actions>
Stop / escalate when: <repeated failure · ambiguity · risk>
Report each run: <cause · change · proof>. Idle -> say so and end.
The four guardrails (non-negotiable — this is the leash)
- Explicit stop condition. Every loop ends: target met, N runs with no progress, or nothing-to-do. A loop with no stop is a runaway.
- The Never-gate. Name the irreversible actions the job must NOT take and must surface instead: push to a shared branch, run migrations, touch
.env/secrets, deploy, delete, send. The kernel already gates these (assess() -> Ask/Block), but state them in the prompt so the agent self-stops before even proposing them.
- Per-run report. Cause · what changed · proof (green/red, counts, file:line). No silent runs.
- Give up and flag. After 3 attempts on the same failure, stop and surface it — don't thrash.
Solo cut
Solo (no PRs, no review gate): a loop works your branch / working tree directly, so the leash is the Never-gate + commit, don't push. git_commit on the current branch is reversible; git_push to a shared branch is not — keep it behind the gate.
Attribution
Adapted from Boris Cherny, Why Coding Is Solved (Anthropic, 24:33), §2 "Loops & Routines", via the build-catalog extraction. Mapped to Vanta's primitives (vanta run / vanta schedule --cron / vanta cron).
