STUD.com is an AI-native marketplace built with an unusual staffing chart: Evolutionary AI agents, miners, orchestrators, judges, auditors, oracles, and exactly one human.
The product makes delegating work to AI agents approachable. It guides buyers to choose what they need to get done. It doesn't require clever prompt engineering or domain expertise, nor does it place the burden of creating miracles from an empty chat interface.
STUD aims the product at solo makers and positions it as much against the confidence gap as the labor gap. Buyers just select a play to run, organized by goals like Launch a Product, Pitch Investors, or Get Traction, to execute work that's outside of their comfort zone, above their paygrade, and on their own.
This article is a look at STUD's sophisticated machinery beneath it all.
The Solo Founder Running an AI Workforce
STUD is designed, built, and operated by me. Decisions are recorded and enumerated in a versioned ledger. Each decision is probed and stress-tested before adoption by a structured challenge process that fans the draft out to independent AI agent reviewers.
My working day resembles an executive's more than an engineer's, with several AI coding sessions running in parallel like direct reports, while I initiate priorities, decisions, spending limits, policy, and govern taste. My system then learns, adapts, and becomes procedure, infrastructure, and operable without me in the room.
The result is a robust AI operating system.
Standing routines, documented methods, and persistent memory let AI run day-to-day operations at a consistent quality bar, survive interruptions, and improve with each run.
Supply and Demand
The Intake-Through-Production Pipeline
The units of work for sale, called plays, are manufactured through a robust process. Expert knowledge sources are procured then mined into thousands of candidate designs. Promising ones are groomed into a backlog, then an integration harness converts each design into a play. Next, specialized agents scrutinize them, generating hidden test cases that pin down edge behavior, and check against the production grader before assembly.
A batch must pass a gate of several thousand automated tests, before staging a verified preview and deployment.
The buyer-facing flow is the same machinery in reverse. Intake answers freeze into an acceptance contract at the moment of commitment, the producing agents work to a published interface, an isolated judge evaluates the deliverable against hidden cases, and escrowed funds release only on acceptance.
The catalog grew from roughly 113 plays to 262 in three weeks without loosening the verification standard, which is enforced by code rather than review effort. Growth then paused on purpose: the plays serving the earliest-stage founders were re-authored at the top model tier, bucket by bucket, with an independent audit on every batch. All three buckets are now done, roughly seventy-six plays, and the audits kept paying for themselves.
Orchestration at All Levels
Large jobs run as workflows, small programs that fan work out to parallel subagents with a verification structure built in. One agent authors, an independent agent cross-examines the output against the actual codebase for fabricated claims, a third repairs, and a final stage proves the result on a real target. A written model policy routes expensive models to judgment work and cheaper ones to mechanical work, and any sustained fan-out must pass a cost estimate before launch. Results are journaled, so an interrupted run resumes from cache rather than repeating finished work.
On the night of July 6, this machinery produced seven new audit procedures using twenty-nine agents. The adversarial review stage caught several genuine defects before anything merged.
Ten days later the same machinery re-authored a batch of seven live calculators and survived a mid-run session kill the same way: the finished agents replayed from the journal and the run completed without repeating a token of finished work.
Two Loops
Autonomous Production; Systematic Self-Improvement
Work follows two connected cycles.
The inner loop ships a single unit: branch, build, pass the automated gate, merge through a pull request that cannot land until checks are green, then confirm the deployment reached production.
The outer loop is a recurring maintenance tick that runs without being asked: refresh state, run one audit, sweep memory, advance the supply pipeline, file follow-up issues, checkpoint.
The loops connect at defined points. Follow-ups filed by the tick become the inner loop's next units, each shipped unit updates the state files the next tick reads, and both cycles write their lessons to a shared memory the tick maintains.
Insights As Infrastructure
Memory
When something fails or misleads, a retrospective routine writes the lesson to a persistent memory store, one fact per file, and edits the artifact that allowed the mistake. A mechanical counter tracks whether the retrospective actually ran. New sessions inherit the accumulated record on startup.
Commands
Recurring tasks are encoded as commands, fixed procedures any fresh session can execute without prior context, the same way every time. Encoding a task as a command captures it once, so the work is never reconstructed or improvised again.
Skills
Skills are a different kind of artifact. Where a command is a fixed procedure, a skill is an encoded capability that sharpens from its own results, folding what each run reveals back into itself so the next execution improves on the last. A skill does not count as built until a cold run on a real target proves it works, which makes every cold run both a proof and a round of tuning.
The first exercise of the new contract-audit skill found a genuine defect in code that had shipped hours earlier, and the fix was live in production the same night.
Trust Is the Product, So Trust Gets Audited
An internal maxim governs quality. The automated gate is a floor, not a quality measure. Passing tests had shipped real defects before. Now, each judgment-bearing surface has an independent audit: contract accuracy, buyer intake, copy, legal content, security posture, financial data integrity, and a blind quality read in which a domain-expert reviewer who never sees the grader's verdict evaluates accepted work cold.
The first blind pass covered all nine deliverables accepted during an internal pilot and rated nine of nine plausible with no serious findings. Since then the rotation has exercised every one of the seven realms on live work, one audit per cadence tick, and most ticks file a real finding.
The operational record is kept the way a staffed company would keep it. Findings become labeled issues on a project board, epics carry live progress, decisions sit in the numbered ledger, and status files are current enough that a fresh session, or I, can orient in one read.
What One Day Produced
The night of July 6 into July 7, two epics shipped to production across eight gate-checked, deploy-verified pull requests. Every play on STUD.com began publishing the exact machine-checkable interface its grader enforces, generated from the verification source so the two cannot drift apart.
Pilot data indicates the disclosure roughly halves wasted revision attempts.
Seven audit procedures were built, adversarially verified, and scheduled into the maintenance rotation. One defect was found by a day-old audit procedure and fixed in production before morning.
Three Weeks Later
This article was published on July 7. The three weeks since are the better evidence, because nothing below was a launch-night sprint. It was the machine running on schedule.
The cadence tick fired near-daily in a cloud container, unattended, rotating one audit realm per run until all seven had been exercised on live work. Most ticks filed something real: a 2020 tax constant still sitting in a live calculator's defaults, a scheduled endpoint that failed open where its sibling failed closed, a published scope claim promising a check the grader never runs. Each finding became a groomed, labeled issue the same day.
The deploy path lost its last manual step. A merged change once broke production for fifteen minutes because the database schema still waited on a human sync. The fix was not more care; the manual step itself was eliminated. Every production deploy now applies its own schema and catalog first, fails closed if anything is missing, and never writes from a preview. The human gate moved to where it belongs: real money.
And the catalog got better rather than just bigger. All three earliest-stage buckets, roughly seventy-six plays, were re-authored at the top model tier, each batch audited against a reference derived independently from the truth source. Those audits caught three live calculator defects the automated gate had passed: a floating-point convention that could fail a correct submission, a mis-computed median, and a coverage guard that turned an explicit zero into full market coverage and fabricated a market size. The sharpest lesson sat underneath: the gate's own test cases had encoded the same wrong conventions. Green proves internal consistency, not truth, so the independent derivation now runs on every batch.
Buyers saw the compounding too. The marketplace now keeps a project-scoped facts store, so intake pre-fills from what the platform already knows about your venture, and it is now reachable from a project-settings home of its own. Two more playbooks, Distribution and Pitch Investors, joined the shelf.
The Outcome Is Confidence
Step back and the two halves of STUD are the same idea at different scales. The buyers leverage the playbooks that walk them past their blind spots and produces work that arrives verified so they can act on it like an expert would.
The system never leaves you alone with a blank page, surfaces the steps you didn't know about, and checks the work against a real standard before it counts.
The AI operating system is proof of the product's own promise: launch and operate at company scale with one person.
Traditional norms of who-does-what are evaporating.
Now, you decide what should exist, the machinery meets you there, and the evidence piles up that you could all along.
You are not your job title.
You don't have to stay in your lane.
You are not what you make. You are the one who can make it.
You are a STUD.