# This company runs on agents.

We do not hire people for the operating roles. We build the agents that hold them, and we run our products on that fleet.

> Newest snapshot dated 2026-08-19. Figures are committed JSON rendered on the server, not live telemetry.

## 01 — The fleet

Not a free-roaming loop. Each one is a function a durable workflow invokes for a judgment-heavy sub-task: a system prompt, a curated tool subset, an output contract, a spend cap, and a defined escalation.

### Domain (16)

- `sourcing` — Brief to ranked candidates
- `vetting` — Fit score and brand-safety flags
- `outreach_writer` — Grounded personalised outreach
- `conversation` — Classify and answer replies
- `conversation_responder` — Draft the reply that goes out
- `logistics` — Create and track shipments
- `content_verify` — Find the post, match the brief
- `analyst` — Compile and narrate the report
- `research` — Site to sales analysis
- `intake` — One line to a valid brief
- `lead_outreach_writer` — Outbound to brands
- `payment_mandate` — Authorise spend against a mandate
- `compliance` — Disclosure and policy checks
- `creative` — Brief-aligned creative direction
- `a11y` — Accessible copy and structure
- `customer_success` — Activation and retention

### Meta (3)

- `coordinator` — Route work across the fleet
- `critic` — Score before anything ships
- `optimizer` — Tune what the critic scored

### Watchdog (3)

- `anomaly_watch` — Flag runs that drift
- `cost_watch` — Stop spend before the cap
- `security_watch` — Watch the trust boundary

## 02 — The replacement ledger

The left column is quoted verbatim from our internal agent roster, written before this website existed. We did not rewrite them to read better. They read like someone's actual job because they were — ours.

| before · human | after · agent |
| --- | --- |
| the human clicking Search + Step 2 | `sourcing` — Brief in, ranked candidates out, de-duped against the blacklist and every prior campaign. |
| the human eyeballing Step 2's table | `vetting` — Pulls each profile, computes engagement and average views, scores fit, raises brand-safety flags. |
| "AI 작성" button + the human reviewing | `outreach_writer` — Writes multiple angles, runs a four-judge tournament, checks spam score, returns one draft. |
| the human reading the Replies tab | `conversation` — Classifies every reply, extracts the address or the rate, drafts the response or escalates. |
| the human in Step 4 | `logistics` — Creates the shipment, watches tracking, handles the exceptions. |
| the human in Step 5 | `content_verify` — Polls for the post, matches it against the campaign brief, computes what it did. |
| the human in Step 6 | `analyst` — Compiles the report and writes the narrative around it. |
| the 6-tab campaign-create form | `intake` — A short conversation that turns a one-line ask into a valid campaign brief. |

…and 14 more agents doing work that had no human predecessor at all.

## 03 — How the work moves

Work moves through named stages, and every irreversible action sits behind a named gate. Gates ship on by default. An operator moves the whole workspace between three levels — and the level, not the mood of a model, decides what happens without a human.

- Stages: `overview` → `sourcing` → `outreach` → `shipping` → `content_review` → `performance`
- Gates: `approveShortlist`, `approveOutreachSend`, `approveReplyResponse`, `approveShipment` (required), `approveStageAdvance`, `approveContent` (required), `approveBudget` (required)
- Autonomy levels: `copilot`, `checkpointed` (default), `autonomous`

## 04 — The operating layer

The same piece of work, priced two ways. The left column is what an outside operator charges to run it, plus the hours it takes them. Ours is metered agent and infrastructure cost, plus the hours a person spends at the gates.

> The management and operations layer only. The spend that passes through to third parties is identical either way and is excluded — including it would make the difference look larger than it is.

- Outsourced to an operator: $2,400 · 45 human hours (20% management fee on $12,000 of pass-through spend)
- Run by the fleet: $7.4 · 2 human hours
- excluded from this comparison: $12,000 — Pass-through spend — identical on both sides
- One engagement of the size a small team would outsource.

## 05 — How we staff

**We don't hire people. We build agents.**

That is a statement about where headcount goes, not about who is accountable. Two people run this company, and the boundary between what they decide and what the fleet executes is written down, versioned, and enforced in code.

### What stays with a person

- Anything irreversible: releasing spend, signing a contract, final sign-off on published work.
- Setting the autonomy level, and moving it.
- Deciding what the company builds next.
- Everything on this page. Two names, at the bottom.

Gates a person can never delegate: `approveShipment`, `approveContent`, `approveBudget`

### What the fleet holds

- The operating roles a company this size would otherwise hire for.
- The work that runs on a schedule, and the work nobody wants to do twice.
- The checks on its own output — scoring, cost, drift, and the trust boundary.

## 06 — The company itself

Merging to main is the release — no human runs a deploy command. Deploys to the agent service land as a zero-traffic canary and promoting one stays a deliberate human decision: the same gate-shaped governance the products are built on.

- verify-build (lint, build, type-check)
- TypeScript unit suite
- offline Python suite + golden-eval holdout gate
- submission artifact suite
- checksum-pinned secret scan

- Python 3,000 · TypeScript 596

This website is in that loop too. The figures on this page are a JSON snapshot committed to a public repository and rendered on the server — nothing here is fetched after the page loads.

## 07 — What we operate

The fleet is not the product. It is how the products get run.

- [Social Seeding](https://socialseed.ing) — Creator campaigns, run end to end by the fleet. A brief goes in; sourcing, outreach, replies, shipping and verification come out.
- [kbeauty.market](https://kbeauty.market) — A buying service dressed as a shop: pick from a live catalogue, we buy at Seoul retail and ship one consolidated box. Checkout is not connected yet — browsing works, paying does not.
- [Teslam](https://teslam.io) — Drive-to-earn for Tesla owners in Korea. The unit economics are public on the site; the settlement pipeline is a design, not a deployment. An independent project, not affiliated with Tesla, Inc.

## 08 — The lab

The fleet frees the calendar, and the calendar fills with builds. Everything below is verifiable without taking our word for it — a live site, a public listing, a demo video, or a judged result.

- **Satellite** — The live satellite catalogue on an interactive globe — close approaches, re-entries, space weather — briefed by agents in manual, assist and autopilot modes. EN·KO·JA. (live: https://satellite.agentba.se)
- **Travel** — An AI travel planner that plans routes and weather together, and refuses to invent what it cannot verify — no fabricated fares, no imaginary flight numbers. Booking hands off to the real engines. (live: https://travel.agentba.se)
- **Somm.dev** — Six AI sommeliers read a public repository in parallel — structure, quality, security, invention — and return one evaluation with its reasoning attached. Any public repo, no sign-in. (live: https://somm.dev) · (Devpost: https://devpost.com/software/somm-dev)
- **vibeDeploy** — One click runs idea discovery, a live board ranks the candidates, and the chosen app is built and deployed. Took 1st place at DigitalOcean's Gradient AI Hackathon. (video: https://youtu.be/VHw-FdvcF2A) · (Devpost: https://devpost.com/software/vibedeploy-zero-prompt-ai-app-factory-on-digitalocean)
- **VibeCat** — A macOS desktop companion that watches the screen with you and suggests before you ask. Honorable Mention at Google Cloud's Gemini Live Agent Challenge. (video: https://youtu.be/j1zzfoDr7qA) · (Devpost: https://devpost.com/software/vibecat)
- **Glasshat** — An audit layer for AI judging: a six-perspective panel scores with evidence, then audits its own over-confidence and pulls the scores back — every step a visible trace. (video: https://youtu.be/A9diFeRDybo) · (Devpost: https://devpost.com/software/glasshat)
- **Preview Forge** — A Claude Code plugin that turns one line into twenty-six mockups, lets a person pick one, and drives 144 agents to a frozen full-stack app. (video: https://www.youtube.com/watch?v=_xHL8SZqfyI) · (site: https://two-weeks-team.github.io/PreviewForgeForClaudeCode/)
- **GitLab Atlas** — One command turns a GitLab repository into architecture documents that then defend themselves: every merge request is checked for drift, and every merge regenerates the baseline. (Devpost: https://devpost.com/software/gitlab-atlas-code-to-architecture-in-one-command)
- **Memex** — Local memory for coding agents — session history as navigable spatial memory, no LLM at runtime, everything on your machine. Its embedding fix was merged upstream into fastembed-rs. (live: https://memex.quest)
- **He was Socrates** — A fullscreen Socratic bust that listens and only asks back, fully on device — zero network entitlements, first token in 192 ms median, the benchmark committed beside the code. Built for the Gemma 4 Good hackathon. (live: https://he-was-socrates.vercel.app)
- **Find Your Bible Character** — Twenty-four questions read four spiritual tendencies back as one of twelve biblical characters, with the model running on the phone. It is on the App Store, and it deliberately gives no score and no ranking. (App Store: https://apps.apple.com/us/app/find-your-bible-character/id6790217296)
- **vibe-mod** — Write a moderation rule in English; it compiles once into a deterministic rule with shadow mode and a 30-day undo. Live in the Reddit App Directory. (listing: https://developers.reddit.com/apps/vibe-mod) · (site: https://two-weeks-team.github.io/vibe-mod/)

Also on the bench: Fairthon — the six-hat evaluation system Glasshat grew out of, scoring pitch decks and repositories against one rubric; VibeMeeting — a macOS overlay that translates the other side of an English call and drafts three answers you can say yourself; ClaudeSync — keeps AI coding environments in sync across Macs; vibeVoice — the text-to-audio dashboard our demo narrations are made with; openClawWorld — a 2D world where people and agents share the map; Memoed on your life — an evidence-first iPhone app that finds what changed in everyday memories

## 09 — The competition record

Public competitions are the cheapest neutral benchmark a two-person company can buy: outside judges, fixed deadlines, published winner lists. We enter with the fleet and publish the whole column — the ribbons and the losses alike.

| date | event | project | result |
| --- | --- | --- | --- |
| 2026-02 | Gemini 3 Hackathon (Google) | Somm.dev | No award |
| 2026-02 | PlayMCP “Player 10” (Kakao) | kidsafe-mcp | Selected |
| 2026-03 | Gradient AI Hackathon (DigitalOcean) | vibeDeploy | 1st place, $8,000 |
| 2026-03 | Gemini Live Agent Challenge (Google Cloud) | VibeCat | Honorable Mention, $2,000 |
| 2026-03 | GitLab AI Hackathon (GitLab) | GitLab Atlas | No award |
| 2026-04 | Built with Opus 4.7 (Cerebral Valley × Anthropic) | Preview Forge | No award |
| 2026-05 | Mod Tools Hackathon (Reddit) | vibe-mod | No award |
| 2026-06 | Vector Space Day (Qdrant) | Memex | No award |
| 2026-06 | Rapid Agent Hackathon · Arize track (Google Cloud) | Glasshat | No award |
| 2026-06 | AI Agents Challenge · Track 3 (Google for Startups) | SocialSeed.ing | No award |
| 2026-07 | Games with a Hook (Reddit) | Pipuzzle | No award |
| 2026-07 | Global AI Hackathon Series with Qwen Cloud (Alibaba) | Circle Take | Judging, announced 2026-08-21 |

- kidsafe-mcp: Program selection. No public roster page exists, so this row rests on our own records.
- vibe-mod: No award; the app itself passed review and is live in the Reddit App Directory.

The last row is still being judged — the result lands on Aug 21, 2026, and this table will carry it either way.

Judged once: CMUX × AIM Intelligence Hackathon, Seoul, April 2026 — our own record, not a published roster.
And two internal hackathons of our own, scored against the same kind of rubric.

## 10 — What we can hold for you

Not a promise of outcomes — a choice of engagement shapes. Each one points at the part of this page that backs it.

- **Run** — Creator campaigns, operated end to end: a brief goes in; sourcing, outreach, replies, shipping and verification come out. This is Social Seeding, already operating — section 07.
- **Build** — An operating seat in your company, replaced the way ours were: a named roster, output contracts, spend caps, gates that default to on. Sections 01–06 are the spec.
- **Prove** — Two weeks to a working system: fixed scope, a fixed fortnight, and at the end something you can operate. The cadence the competition record was built on — section 09.

## 11 — Where we operate

- Operating: Seoul, Tokyo, Ulaanbaatar, Mexico City, Hanoi, Bangkok, San Francisco, New York, Paris, Milan

## 12 — Intake

**One line is enough.**

The fleet has an agent named intake, and its whole job is turning one line into a valid brief. Send a line about your work — the thing that runs on a schedule, the thing nobody wants to do twice. Both founders read every line.

No form? The addresses in the footer work the same.

## Grants & programs / Built with

- Grants & programs: ElevenLabs Grants (https://elevenlabs.io/startup-grants)
- Built with: Google Cloud, Google Gemini, OpenAI, Vercel, Next.js, MongoDB, Go, Tailwind CSS

---

- CEO Sanggen Chang — sangguen@agentba.se
- CTO Sejun Kim — sejun@agentba.se

https://agentba.se