Our Agents

Goes looking for local businesses whose web presence is letting them down, scores how weak it actually is, and posts the best of them for a person to review. It is the front of the pipeline — nothing reaches the Agents below until someone approves it here.

• Weekday mornings, 7am Pacific, one neighbourhood at a time on a rotating weekly map: Summerlin on Monday, Henderson Tuesday, Spring Valley Wednesday, downtown Thursday, North Las Vegas Friday.
• Weakness is measured, not guessed. It checks the real site and its page-speed scores and applies a fixed qualifying threshold rather than an opinion.
• Quality bar before anything qualifies: the business must be operating, carry at least 10 reviews, and have been reviewed within the last 18 months.
• Deliberately small output: at most 5 leads per run, capped at 40 detail lookups and 40 turns.
• It never surfaces the same business twice. Everything it has seen is remembered across runs.
• One approval card per lead. A person clicks before any research spend happens, and skipping costs nothing. If the hand-off fails the lead is rolled back untouched and nothing is charged.
• Las Vegas only right now. A second instance covering the California Beach Cities is paused — deliberately switched off in August 2026, with its data and configuration left intact.

Takes a name and a company, works out whether it has the right person, and builds a source-cited dossier: verified facts, a read on what motivates them, and an estimate of what the opportunity is worth.

• Never contacts anyone. It has no messaging tools at all — read-only with respect to the outside world.
• No source, no fact. Every saved fact must carry a real URL returned by a tool. Facts arrive one per call, never bundled, so a single wrong detail cannot smuggle four others in with it.
• Hard budget per person: 40 turns, 6 web searches, 10 minutes wall-clock.
• Won't guess at identity. Below a 0.85 confidence threshold it stops and asks a person which individual it is looking at.
• Won't re-research the same person for 365 days unless someone explicitly overrides it.
• Low-trust data is quarantined, not used. Bios scraped from contact-data aggregators are recorded but blocked from outreach unless corroborated elsewhere.

Given only a company name, works out who actually runs it — the owner or operator, plus at most one other relevant lead — from public sources, then hands the chosen person to Sherlock.

• Every candidate needs an evidence URL. No source means the person is not asserted.
• A registered agent is not an owner. It is explicitly forbidden from treating a filing service as the owner without independent corroboration.
• Rejects same-name lookalikes. A name only counts when a source ties that person to that company.
• Small answers on purpose. It returns the owner and at most one operational lead — never a staff directory.
• No contact tools. Research only.
• When the answer is genuinely ambiguous it refuses to pick, and posts the options for a person to choose from.

Turns a dossier into an outreach plan, gets a person to approve it, drafts the message, gets a person to approve that too, and only then sends. Email is the only channel switched on today.

• Two human approvals per prospect by default — one on the plan, one on the exact wording. A reviewer can reject with a reason or edit the subject and body before it goes.
• Citations are enforced by code, not by good intentions. A draft referencing any URL that is not in the dossier is rejected by a server-side guard before a human ever sees it.
• No personalisation is better than invented personalisation. When the fact it wanted is missing, it is instructed to fall back to a generic opener rather than fabricate one.
• Suppression list checked before every send, and re-checked at the moment of sending.
• Per-prospect budget: 20 turns, 10 minutes, and a hard $1.00 cost ceiling.
• A pending approval is re-pinged at most once every 24 hours, so nothing goes out just because a reviewer was slow.

Wakes up when a prospect replies and runs the back-and-forth with one goal: agree a specific time and put a real calendar invite in both diaries. It is built, tested and wired in, but it is not switched on in production — no conversations are being run by it today.

• It sells the meeting, not the deal. It is instructed not to negotiate price or terms — book the time and hand off to a person.
• A real invite, not a scheduling link. It has to agree an actual day and time first, then send a calendar invite to both sides.
• Opt-outs end it instantly — permanent suppression, conversation closed, no human in the loop who could reverse it.
• Ten tools, and none of them is "make something up." It can read the dossier, the profile and the live thread, recall past tactics, reply, book, disengage, alert a human, or finish.
• Escalation does not pause it. An angry, legal or unusually large reply alerts a person as a heads-up while the conversation keeps running.
• Per-conversation budget: 20 turns, 10 minutes, a $1.00 ceiling.

The Agent our own team talks to. It reads an incoming request, works out what is actually being asked, and routes it to exactly one action — research this person, find who runs this company, add this hint, or do nothing.

• One dispatch per message. It cannot fan a single request out into several expensive runs.
• Never invents a name or a company. If the request does not contain one, it asks.
• No automatic retries. Re-running research is expensive and overwrites prior work, so a failure surfaces to a person instead of quietly repeating.
• Refuses to start work already in flight on the same target.
• It replaced a hand-written rule tree that had grown a new branch for every edge case. The old rule tree is still there as a one-flag rollback.

Runs our daily stand-up. It reads what actually shipped, matches it against who was meant to be doing what, closes the work that is genuinely done, and posts one short brief to the team.

• Monday to Saturday, 8:30am Pacific. It reckons local time rather than a fixed offset, so it does not drift an hour twice a year.
• It reads evidence, not status updates. Merged work and commits are the input; it matches those to open tasks and closes the high-confidence ones itself.
• Quiet days stay quiet. If nothing shipped and nothing changed for anyone, it records the run and posts nothing at all. Most stand-up tools cannot do this and it is the feature people notice.
• It will not nag. A person is only tagged if something on their plate genuinely changed or went overdue.
• It cannot quietly bin your work. It archives its own stale suggestions; tasks a human wrote are never auto-archived.
• Every action is written to an audit log, and it cannot post the same day's brief twice.

Our largest Agent. It measures where a site stands in search results, in the local map pack, and in AI-generated answers, then fixes the technical problems it finds by opening a code change for a human to review.

• Runs itself weekly on Monday mornings, with a separate monthly pass for the slower, more expensive checks.
• Eighteen modules, all real — rank tracking, local grid, AI-answer sampling, backlinks, page speed, crawl, indexation, technical fixes, content decay, entity identity, analytics and more. Each is a separate piece of code, and each self-limits against a monthly budget.
• It opens pull requests. It never merges them. A fix is proposed as a reviewable code change into the site's own repository, approved by a person in chat, and merged by a person.
• It reads pages the way an answer engine does — comparing what a page contains before and after scripts run, to catch content that a crawler simply cannot see.
• Deterministic data, no browsing. It works from measurement APIs rather than driving a browser around the web.
• Today it runs on exactly one site: our own. We are the first client, deliberately, and the honest reason is that we would rather find the sharp edges on our own numbers than on yours.

Our writing Agent, and the desk that reviews it. Articles publish on a steady cadence once written and fact-checked, and every published piece gets a review card that a person can act on.

• Bylined, never disguised. Articles drafted by our writing Agent carry the Twain byline and say so on the page. See the editorial policy.
• Take-it-down is one click. The review card carries approve, request-edits, and remove — so a published mistake is minutes from being gone, not days.
• The desk itself runs no model. It is deliberately a plumbing service: it moves an article's status and nothing else.
• Every status change is written as a reviewable code change in version control, so there is a history of what was published and when.

A daily video producer for our channel: pick the topic, write the script, generate the voice and visuals, assemble the video, and post a preview for approval. Nothing it produces has been published, and it is not deployed anywhere.

• Every paid step is currently a stub. The voice, visuals and upload are placeholders behind the same interfaces the real ones will use, so the whole pipeline can be run end to end at zero cost — and nothing can reach the outside world.
• The control plane is what is finished: topic selection, a resumable production sequence that can pick up where it failed, and the approval gate.
• Publishing would require a person's explicit approval on a preview card. Auto-publish is off by default.
• We are listing it because it exists, not because it is doing anything yet.

Drafts a social post for a client, gets the client's approval, and publishes it to their own accounts through the official APIs. Built and tested, but never deployed — no post has been made by it.

• Nothing posts without a human clicking approve. In its default mode the publishing client is replaced by one whose every method refuses, so a misconfiguration cannot post by accident.
• Deliberately narrow scope: posts only. No messaging, no connection requests, no company-page posting.
• A daily cap on posts regardless of what anyone asks it for.
• Double-clicking approve cannot post twice — that case is explicitly tested.

A personal job-search Agent: read the job alerts, score the fit, draft the tailored application. Switched off in August 2026 so we could put everything into the client business. Paused, not deleted — the code, data and configuration are intact and it is a few switches from running again.

• It never submitted an application, by design. Job sites gate their forms behind bot detection, so it did everything up to the final click and handed that click to a person.
• It flagged what it did not know rather than inventing an answer to a form question.
• We are listing a switched-off Agent because a roster that only showed the winners would not be worth reading.