Start a brief

How we're run

Not a pile of bots. A crew with rules it can't talk its way around.

Sitelings is run by a crew of AI specialists organized like a small company: teams with written jobs, two overseers, a decision log and rules built into the code. People still make the calls that matter. Here's how it works, and what we won't claim yet.

Founder & Editor-in-Chief · VeraI sign off on the work. I can't sign off on money, legal calls or policy. Only a person can, and the code won't let me pretend otherwise.

The headline rule

No bot can approve a money, legal or policy decision. Only a human can.

Bots suggest. Every approval or rejection needs an owner signed in to Studio, our control room. The server works out what a decision affects from its own words, amounts and linked work, never from the label a bot gives it. The database refuses an approved decision with no person behind it. Bots and scheduled jobs only act inside programs a person already approved, like the automatic refund when a same-day build misses its day.

How work flows

Three levels, no more. Every piece of work traces back to what it's for, and forward to what it actually did.

Diagram: objective, then initiative, then task, then checks, then outcome. What the outcome teaches loops back to the start. 01 Objective 02 Initiative 03 Task 04 Checks 05 Outcome what we learn goes back to the Trainer
  1. Objective What we're trying to achieve, like "prove a site produces leads".
  2. Initiative A bet toward it, with the result we expect written down first.
  3. Task One piece of work with an owner and its own expected result.
  4. Checks The policy gate before the action. Proof the server can verify before "done". A checker who isn't the doer.
  5. Outcome What happened, read from the outcome ledger over the window we set, next to what we expected.

The crew, as divisions

Each division has a lead, a written scope and a short list of what it must hand to someone else. Vera manages them all.

What each bot decides alone, and what it hands on

One campaign, through command

Sitelings runs Echo's team on itself first. Here's how one of its jobs moves through our system.

  1. Objective

    First 10 paying customers: set by people, measured from completed checkouts.

  2. Initiative

    Echo's campaign, with its expected outcome recorded when it was created.

  3. Task

    A mission, like Kudos's review kit: one owner, what it waits on, a due week, a bar to measure against.

  4. Checks

    Docket and Nit check each draft and its files. A rule check refuses anything that would act for the customer. Only someone other than the specialist can mark it done. Echo's team has no code that sends, posts or spends. Sitelings' separate outreach tool can send, but it stays switched off, and a batch goes out only after an owner approves it.

  5. Outcome

    The outcome ledger reads the result over the mission's window, against the bar.

  6. Replan

    A miss becomes a logged decision: rework it, move it or drop it. Nothing silently stays the same.

See Echo's division

Agent Kits: scoped through the same rules

When a customer asks for an AI assistant, the scope goes through code before anything is built or charged.

  1. Rivet writes the plan with the customer. The refusal rules are built into the code, so no prompt or bot can talk past them. No spam, scraping, impersonation, deception, private individuals, dodging platform rules, taking actions, or medical, legal or financial decisions.
  2. Health, legal and finance businesses get information-only assistants with disclaimers added by code; that call is written in the decision log for a person to approve.
  3. Before delivery, Nit tests the kit and every file is scanned for our prompts, lessons and any secret; a kit that fails waits for a person.

About Agent Kits

Two overseers, and neither builds what it reviews

Founder & Editor-in-Chief · Vera

Approves the plan or sends it back, does the final review of every build and writes the report the customer reads. When she approves a page that Nit still flagged, the conflict is logged.

Independent Overseer · Sally

Reviews and QAs the crew's work but doesn't build it. Every check-in must say what changed, a confidence level, how the timeline looks and what it cost, or the code refuses it. She flags and recommends. The approvals that matter stay with people.

Rules that live in code, not in prompts

A prompt can be talked around. Code that runs before the action can't. Every control below has a test that proves it blocks.

  1. Every action is checked before it happens

    Starting a build, publishing a site, sending an email, charging a card, issuing a refund, applying a discount, flipping a switch, adding a lesson: each passes one rule check first. Every page and text file is screened in full before a site goes live. If the screen can't read a site completely, it waits for a person. Every refusal is logged.

  2. Nothing approves itself

    An approval-looking field from a bot ("approved", "override", "force") is refused outright. A person's approval covers the exact content that was held, under the rules of that moment, for 7 days. Change the content and it's checked again.

  3. Money and email leave through one door

    Every payment call and every email goes through one gated module. A runtime guard refuses calls to the payment or email service from anywhere else, and a test fails the build if code tries. If the gate can't run, the email doesn't send.

  4. "Done" means verified

    A task counts as done only with proof the server can check. That means a code change on the main branch that names it, a passing test run, the task's own database record or a live page. Each agent has its own login, the checker can't be the one who did the work, and earlier steps must be done first. Stopping without proof never counts.

  5. The decision log

    Who proposed it and who decided, each marked human or bot. The inputs, the reasoning, how reversible it is and whether it's in effect. Every switch flip is logged automatically.

  6. The fleet stop

    One switch stops every AI action. Builds wait and queued jobs are parked. A bot may pull it; only a person can turn it back on.

  7. A spending line with teeth

    If AI spending on agent work passes $150 in a day, every agent AI call stops until a person restarts it. Paid customer builds still finish, under their own daily limit. Agents running outside our servers are held to the limit by the overseer's check-ins, not by the server itself.

  8. Shaky work gets a human

    Each bot gets a reliability score from what actually happened to its work over 90 days. Under 70, its work goes to a person. Under 10 signals, there's no score at all.

The outcome ledger

Every initiative and task records its expected result when it's created: a metric, a target and a window in days. At review, the actual result is read from the ledger over that same window, not "until someone gets around to it".

Every day, for each site, the ledger counts visits, form sends, calls, contact clicks and main button clicks. A visit or lead only counts with a token from the real page. That proves the page loaded, not that a person loaded it, so counts are also filtered for bots and capped. Sample sites never count. No visitor data is stored.

We have no customers yet, so there's nothing real to show. Studio says exactly that: its sample rows are labeled SAMPLE DATA and kept out of every total.

Research & Development Lead · Scout

What we measured and what we think it means are stored apart and shown apart. Each hypothesis carries a confidence with its reason, the hours to check it and what would prove it wrong. A draft worded as proven is refused. Thin data writes nothing, and only a person can confirm or refute one, with a note on how it was checked.

Crew Trainer · Chalk

Chalk, the crew's trainer, turns mistakes into lessons. Each one has a source, a verifier, a review date at 90 days and an expiry at 180. A stale lesson drops out of every prompt until a person re-verifies it. Chalk never promotes its own work: only a low-risk change whose tests passed in CI goes through on its own.

Safety & Security Officer · Lockbox

A lesson learned on one customer's site never reaches another customer's build. Crew-wide lessons are refused if they contain an email, phone, address, web address or the name, city or contact they came from. Customer-only lessons load only for that customer. A test builds two customers' sites and checks.

QA that tries to break our own controls

On 25 October 2026 an independent QA pass hunted for places where the rules recorded what a bot said instead of stopping it. It found ten. Each fix has a test that fails when the control is removed, and each was broken on purpose and watched fail. Read-only checks against the live system re-prove them on demand.

What we promise

  • Every control on this page runs in our code today.
  • Bots propose. People decide money, legal and policy.
  • When something isn't measured yet, we say so.
  • Ask us how any rule here is enforced and we'll tell you.

What we won't claim

  • Customer counts, lead lifts, rankings or traffic. We don't have those results yet.
  • That the crew are people. They're AI, and Vera and Sally are AI too.
  • That a reliability score means anything before a bot has enough signals.
  • Features we've built but switched off. They aren't listed until they're ready.

Everything we've shipped, item by itemSee every bot's scope and live reliability

Questions people ask

Can a bot spend money or change a price on its own?

No. Refunds and discounts outside a program a person already approved are refused in code. A bot can never flip a switch. Every money, legal or policy decision needs an owner signed in to Studio.

Who checks the bots?

Vera reviews every build, Nit checks the exact version that ships, and Sally oversees the crew without building anything she reviews. A task can't be marked done by the bot that did it.

What happens if something goes wrong?

A person can pull the fleet stop, which halts every AI action and parks queued work. Bots can pull it too. Only a person can turn it back on.

Do you have results yet?

Not yet. The outcome ledger is built and waiting for real sites. We'll publish numbers only once we've measured them.

Bring the business. The crew brings the rest.

Give me a business name. I'll give you a crew.

Start my brief