Building Live system · runs weekly

The Editorial Agent.

An editorial agency, run as agents — end to end. It senses demand, plans briefs, researches primary sources, produces pages and diagrams, verifies itself adversarially, publishes through CI, cuts its own distribution creative, and grades the results against targets it registered in advance.

This site writes a lot about accountability infrastructure — versioned constraints, adversarial checks, append-only measurement. The Editorial Agent is that argument running in production on the site itself: the rules live in versioned rubrics, an adversarial gate stands between drafts and the public, and a measurement loop the system cannot edit decides whether the work is working.

Read the build tutorial → ← Back to Building
The Editorial Agent — one loop, eight handoffs, a rulebook and a human at the center THE EDITORIAL AGENT One loop, eight handoffs. RUNS WEEKLY · SELF-GRADED SENSE GA4 + GSC weekly pulse PLAN demand becomes briefs RESEARCH primary sources, cited PRODUCE pages + diagrams VERIFY the adversarial gate PUBLISH CI to the edge DISTRIBUTE carousels + video loops MEASURE snapshots, self-graded VERSIONED RULES the rubric agents answer to HUMAN EDITOR approves · owns the byline Demand becomes a brief; a brief becomes a page; the gate decides; the loop measures itself.

The loop, station by station

  • Sense

    A zero-dependency fetcher pulls GA4 and Search Console — clicks, impressions, CTR, position, query-by-page for the standards library — on a weekly GitHub Actions cron. As of July 2026 it also reads a second demand channel: Ask Evgeny, the site’s own knowledge agent, contributes the questions readers actually asked — what the corpus couldn’t answer, what got a thumbs-down on an answer it did give, and what’s repeat-asked.

  • Plan

    Demand becomes briefs: striking-distance pages, decaying essays, query clusters nobody owns — now weighted by what readers actually asked, not just search behavior. A gap that’s both missing from search results and repeat-asked in the chat outranks either signal alone; a thumbs-down on an existing page becomes a revision brief, not a new one. The monthly review turns the snapshot into a ranked action list.

  • Research

    Each brief gets a research agent that works from primary documents — release notes, standards PDFs, platform docs — and returns a fact base with a citation per claim.

  • Produce

    Builder agents own one file each: the page, its signature SVG, its schema. Parallel waves, no shared-file collisions, structured returns between stages.

  • Verify

    An adversarial gate re-checks the builders: facts against sources, schema validity, house conventions, leak and regression sweeps. It fixes what it finds and reports what it fixed.

  • Publish

    Green gate, one commit, CI to the edge. Sitemap lastmod, llms.txt, and the internal-link mesh update in the same change.

  • Distribute

    Carousel decks and video loops render from the live pages themselves — same pixels the site serves, captured and cut for the feed.

  • Measure

    Snapshots commit to the repo; the review rubric grades results against pre-registered targets at fixed checkpoints. The loop closes where it started.

The gate is the product

Anyone can chain agents. The part that earns trust is the station that assumes the others are wrong. The verify gate re-derives claims from primary sources, validates every schema block, sweeps for leaks and regressions — and fixes what it finds before a human ever reviews. Its first production day caught a mis-attributed study, a wrong fee basis, and a standard listed as draft that had gone final. Those corrections shipped; the errors never did.

The verify gate — agents check agents, with receipts THE VERIFY GATE Agents check agents — with receipts. BUILDER OUTPUT pages · diagrams · claims ADVERSARIAL GATE · facts vs primary sources · schema, links, accessibility · house conventions · leak + regression sweeps SHIP CI build → the edge moderate or worse: fix, rebuild, re-verify CAUGHT IN PRODUCTION — REAL CATCHES A mis-attributed study a headline variance stat credited to the wrong vendor — corrected in three files before publish. A fee basis, wrong “2.5% of media cost” corrected to “of net revenue, vendor-stated” — then canonicalized site-wide. A “draft” that was FINAL a standard’s status corrected to FINAL against the primary document, dated to the day.

What A2A looks like when it ships work

Agent-to-agent is easy to demo and hard to operate. Five mechanics make this loop hold together:

  • Structured handoffs: every agent returns typed output the next stage consumes — a fact base, a build report, a verdict table.
  • File ownership: one writer per file per wave; shared surfaces (indexes, maps, feeds) belong to a single integration agent.
  • Adversarial pairing: the verifier is a different agent with a different job — refute, not confirm.
  • Constitution over prompts: the durable rules live in versioned skill rubrics the agents answer to, not in one-off instructions.
  • Human at the top of the loop: briefs in, approvals out. The byline — and the mistakes — belong to a person.

Running in the open

The measurement half is public plumbing: a weekly pulse every Wednesday, a monthly content review on the 3rd, snapshots committed to the repo, reports filed as GitHub issues, and checkpoint targets written down before the results arrive — so the system is graded against what it promised, not what it got. In its first full production day the loop shipped twelve research-verified reference pages, upgraded eight diagrams to signature grade, and cut a queue of carousel decks and video loops from the live pages — every claim carrying its citation. As of July 2026 the pulse also carries paid telemetry: campaign spend and clicks read through GA4, cost-per-lead measured against the thanks-page conversion rather than the ad platform’s own count, and landing-page performance for the site’s ad campaigns — report-only, by design.

How a content agent earns an ad budget

The newest capability is deliberately incomplete. The agent can now see the site’s Google Ads activity — three rings, only the first of them live:

  • Ring 1 — read (live). Paid performance flows into the weekly pulse through GA4’s ad link: spend, clicks, honest cost-per-lead, and which audiences the site’s own content mints. The pulse reports; it is forbidden to advise — at small budgets, weekly paid numbers are anecdotes.
  • Ring 2 — recommend (next). The monthly review earns the right to propose: creative refreshed from what performs organically, negative keywords from search-term leakage, budget changes argued from a month of evidence. Proposals arrive as pull requests; a human disposes.
  • Ring 3 — act (deliberately withheld). Write access to the ad account has a published bar it has not met: an allowlist of reversible operations, an approval gate on every mutation, an append-only ledger, and a spend cap enforced outside the agent. The same instruments this site’s essays demand of any agent that touches money — applied, first, to our own.

That sequencing is the argument of the whole build: an agent earns each ring the way it earned publishing — months of verifiable reads and good proposals before its hands go anywhere near the controls.

Where it connects

  • Fire the Dashboard, Hire the Agent →

    The tutorial essay — building the measurement half of this loop, step by step.

  • Ask Evgeny →

    The demand-sensing partner: its gaps, feedback, and repeat-asked questions now feed the Sense and Plan stations directly.

  • AI Growth Operator →

    The sibling build: the same bounded-autonomy pattern pointed at SMB growth instead of editorial.

  • Agentic Transformation →

    The operating playbook this system is a working specimen of — pilots to governed workflows.

  • The Standards Library →

    The loop’s primary output surface — reference pages the gate keeps honest.

  • Right Now →

    The weekly cadence in public: the Friday Thought Experiment the loop publishes around.

Describes the system as it runs today: weekly and monthly crons live, gate mechanics as operated, production-day counts as shipped. Agents draft and verify; a human editor approves, publishes, and owns the result. The Ask Evgeny demand-sensing channel (added July 2026) is live code, wired into every fetch — it has not yet completed a full automatic monthly cycle.

Next step

Want this applied to your GTM, data, or product system?

The builds show how I think. The advisory applies the same operator logic to live commercial, technical, and organizational decisions.

See advisory services Take the readiness scorecard