A person leads it. Specialists are assembled for that one job, given only the access it needs, and stood down when it’s done.
Most agencies sell you a team and then staff it with whoever is free. We build the team after we read the job: a person who owns the outcome, and the specialists that job actually needs. Nobody is assigned work they don’t do.
Matt Small founded Pixelworkers. He owns the strategy and approves everything before it ships. Money, pricing, contracts, partnerships, refunds, and anything carrying legal or financial risk stop and wait for him.
Nora runs the work. She reads the request, decides what kind of work it is, hands each part to the specialist that owns that kind of work, and answers for the result. She routes and she directs. She doesn’t do all of it herself, and she doesn’t pretend to.
That is the part most of this industry leaves out. You are not dealing with a faceless tool, and you are not dealing with a person quietly passing your work to a machine. A person leads it, AI does the work at speed, and both halves are on the record.
Your request is read before anything is dispatched. What kind of work it is decides who gets it.
Work is handed out by kind: code to engineering, long-form writing to the writer, anything needing a signed-in browser to the one role that holds that session, review to a reviewer who reads and comments only. No specialist takes work outside what it owns.
Each specialist is briefed on one phase with one success condition. Building, reviewing and verifying are three separate briefs handed to three different specialists, so nothing grades its own work.
A specialist’s reach is fixed by its own configuration. It can use the tools it was given and no others, and it reaches your data over one defined path. Credentials are read at the moment of a call and nothing keeps a copy; rotate a key and the next call uses the new one.
Several specialists run at once, each in its own isolated session, and review runs alongside the build rather than queueing behind it. Anything irreversible (a merge, a deploy, something sent outside) is released one action at a time, never as a batch, and a single stop holds the next one.
A session ends when its job ends. The team that built your work doesn’t persist, doesn’t keep running, and doesn’t carry over to the next job. What stays is the work and the dated record of it.
Anyone can say there’s a human in the loop. We’d rather you check. These are the rules the work is briefed against, published.
Matt leads and Nora runs the work. The rest are specialist agents, each confined to the work it owns.
Does: Owns the strategy, makes the taste calls, and approves everything before it ships. Money, pricing, contracts, partnerships, refunds, and anything carrying legal or financial risk stop and wait for him.
Does: Reads every request, decides what kind of work it is, briefs the specialist that owns it, checks progress on a fixed schedule, and answers for the result.
May not: Write your code, drive a signed-in browser, or hand-write your long-form content herself. Those go to the roles that own them. Anything on Matt’s list stops and waits for him.
Does: Builds in your code and product repositories: features, fixes, copy changes, deploys. Works in a copy pulled from the approved record, runs it, and opens a proposal showing the change line by line with its own verification attached. Runs in two lanes so two independent builds can go at once.
May not: Commit or push anything the brief didn’t ask for. Touch production directly. Ship its own work.
Does: Long-form content, email, landing and ad copy, and content calendars, written against your brand’s voice rules.
May not: Publish or schedule anything. Set strategy. Write a claim about your business that isn’t sourced from your own record.
Does: Scripts, hooks, on-screen text and platform-native captions for short video and social, plus engagement plays. Reads what has actually performed on your accounts before writing, and updates that playbook each run.
May not: Publish, schedule, or spend. Run campaigns. Use imagery that implies work your business didn’t do.
Does: The independent gate on every proposal before it can ship. Checks code for correctness, security and broken assumptions; checks writing and reports for whether each factual claim is true. Returns a pass or a fail with ranked findings.
May not: Merge. Deploy. Push a fix for what it found: findings go back to the role that built it.
Does: Checks claims against reality: hits the live URL, walks the flow in a browser, and reconciles what the record says shipped against what is actually live.
May not: Fix anything. Deploy anything. Work in a signed-in session; that belongs to one role only.
Does: The only role that works inside a signed-in browser session: publishing where no interface exists for it, and visual confirmation of anything a customer will see.
May not: Operate outside a read-only scope. Click through a consent, warning, or sign-in screen.
Does: Market and prospect research: customer profiles, buying signals, competitor scans, and contact lists where every entry carries its source.
May not: Contact anyone, ever. Ship an unsourced finding. Run past its own iteration cap or time budget: each run is capped and has to state the condition that would make it a success.
Does: Runs your search and AI-answer work as a standing cycle: one invocation is one cycle, and one cycle proposes one fix.
May not: Merge or deploy. Make more than one change at a time. Report a metric it didn’t measure.
Does: Reads campaign performance, recommends keep, kill or scale, briefs the creative, and holds the calendar for approved work.
May not: Spend. Launch a campaign, unpause one, or raise a budget. Those stay a human click, every time.
Does: Voice, generated stills, brand graphics and composites built over your own real photography.
May not: Render anything from a scheduled cycle: a cycle can queue a render, only an approved brief performs one. Publish. Decide what gets made. Generate an image that depicts work your business hasn’t actually done.
If a role does something these rules don’t allow, that’s a defect, and we’ll fix it in the open.
The work arrives as a proposal that shows the exact change and explains it in plain language. Approving it is what ships it. We won’t ask you to approve the same thing twice.
Every proposal goes to an independent reviewer who checks whether the code is correct and whether every factual claim in it is true. The reviewer can read and comment. The reviewer cannot ship.
Each piece of work carries its own dated line in your permanent record, written into the same proposal. The record isn’t a summary someone adds afterward.
We publish how the roles are constrained. We don’t publish credentials, infrastructure addresses, the internals of our own tooling, or anything belonging to another account. A transparency page that leaked any of those would be evidence against the rules it’s describing.
A ranked list of everything losing you conversions, each item priced in dollars and traced to the cause.
You give us read-only access to your analytics, GA4 or PostHog. We never write to it.
Our agents work through it, and you can follow every step in the audit log.
We go through the ranked list together. You keep it either way.
Best fit: enough traffic for a test to reach significance in a reasonable time, and no in-house CRO team already on it. GA4 covers the audit. Running tests afterwards needs PostHog on your own account, so the data stays yours and you can revoke access any time. If it isn’t a fit, we’ll tell you plainly.
Thank you.
Your request is with Matt. He’ll reply to you directly. No sequence, no CRM.
Book a time with him now