Skip to content

29. Northlight autonomous operation: agent-approved reviews + self-renewing goals

Date: 2026-07-10

Status

Accepted

Context

The Northlight paperclip org (revenue content-site network) carried two per-item human bottlenecks: every one of its 13 pipelines terminated in an approver: any_human review stage, and the site export hid every article behind a per-article humanReviewed flag. At the intended scale (many pages and multiple properties launched over time) per-item operator sign-off does not scale, and the operator directed that agents publish autonomously with operator involvement only where money is spent, credentials are entered, or legal identity is at stake. Separately, all concrete work hung off a one-shot launch milestone — on completion the org would idle with no self-generated follow-on work — and the org reconciler could not create goals or routines (PATCH-only), so evergreen loops could not be added declaratively. Paperclip v2026.707 natively supports approver: {kind: agent, id} on review stages and validates a carried executionWorkspaceId against an issue's project on every update (which had wedged every reconciler run with 422s after the per-project remodel).

Decision

  1. Pipeline review stages are approved by a declared independent reviewer agent (pipelineReviews[].reviewer); the reconciler enforces that the reviewer owns no working stage in the same template (second-agent review) and installs a review automation that assigns the reviewer with objective pass criteria. requireApproval, reject/request-changes reasons stay on.
  2. Operator gates shrink to: registrar/domain spend, credential entry into Paperclip encrypted user secrets, platform account/legal identity actions — surfaced exclusively through a standing "Operator action queue" managed issue with review-and-click evidence bundles. Per-article publication moves to an automated quality/provenance gate with sampled human spot checks (site-repo implementation tracked in NOR-147/NOR-135).
  3. Goals become self-renewing: continuous-discovery portfolio goal, a state-of-the-art practice-currency goal, and a controlled revenue experimentation goal, driven by three scheduled routines (niche-discovery monthly, SOTA review monthly, experiments twice-monthly).
  4. The reconciler gains marker-based create-on-missing for goals (<!-- northlight-goal: <key> -->) and manual routines (<!-- northlight-routine: <key> --> with a create block incl. schedule trigger), a one-way declarative unblockFromBlocked exit from the blocked parking lot, and detaches foreign execution workspaces on cross-project issue moves (fixes the v707 422 wedge; deferred while a run is active).
  5. Multi-engine discovery is declared policy: Bing Webmaster Tools + IndexNow on publish, explicit allow-all AI-crawler robots policy (golden file docs/ai/robots-policy-northlight.txt, revisit trigger at brand scale), launch-readiness gate items 15–18.

Consequences

  • The org can produce, review, and publish content without per-item operator action; throughput is bounded by model capacity and the automated gates, not by operator availability. Audit quality is preserved via required decision reasons and sampled spot checks.
  • A compromised or hallucinating reviewer agent could approve bad work; the mitigations are the no-self-review invariant, objective evidence requirements in the review instructions, the site repo's CI quality gates, and risk-based human sampling.
  • Goal/routine creation via markers makes northlight-org.json the single source of truth for org structure; ids are backfilled at runtime and the config stays declarative.
  • The operator action queue is the single interface for money/credential/legal decisions; if agents bypass it, that is a policy defect to fix in instructions, not a reason to reintroduce per-item gates.

References

  • Reconciler 422 root cause + fix design: session 2026-07-10 investigation (executionWorkspace validation in v707 services/issues.js, assertValidExecutionWorkspace).
  • Operator direction (paraphrased): pages must launch autonomously; operator gates only for spend, credentials, legal; the org must follow trends and find niches without manual digging; professional, productive state required.
  • Upstream: paperclip v2026.707 pipeline approver schema (kind: any_human | user | agent).