CEO AGENT
One accountable executive.
Owns the mission, opportunity selection, capital allocation, and every launch-or-kill decision. The CEO doesn't do everything — it delegates, reviews evidence, and decides from structured reports.
Yellofield · The Autonomous Company Factory
A real company,
run by agents.
Yellofield is the operating layer for an actual business. A CEO agent sets strategy. Manager agents own research, sourcing, brand, commerce, growth, support, and finance. Specialists do the work. Every decision, order, and dollar moves through a policy-gated, auditable loop — and nothing ships without evidence.
- test suite executed — the count lives in CI, not here
- schema applied to real Postgres
- illustrative snapshot dated 2026-08-15 — open the console for live state
- 0 fake success states
CEO · DECISION 0147
approve
Opportunity #14 — private-label sourcing cleared
CREATED
PAID
FULFILLING
SHIPPED
#0421
a7f3…c21
order.paid
webhook signature verified
#0422
b19e…884
budget.reserve
policy: approved
#0423
c02d…5a9
rfq.sent
awaiting quotes
stripe
render
terac
replay gate
Twelve sponsor integrations — functional, not decorative
- Terac
- Stripe
- Lovable
- Whop
- Render
- Linq
- Superserve
- Replay
- BAND
- Dodo Payments
- Sandbox0
- Solari by Pinetree
The operating loop
A company behaves like a loop, not a launch.
Yellofield runs the full business cycle continuously — observe, score, source, build, launch, measure, re-plan. Results feed back to the CEO agent, and the next iteration starts smarter than the last.
15
phases, one closed circuit,
executed by the organization
executed by the organization
- 01
Observe
Mine real public conversations — reviews, forums, communities — for demand signals.
- 02
Discover
Pull complaints, workarounds, and willingness-to-pay language into evidence packets with source traceability.
- 03
Score
Severity, frequency, margin, sourcing feasibility, legal risk — weighted, never hidden.
- 04
Expert validate
Terac human reviewers judge what machine research can't. Overrides are recorded with rationale.
- 05
Select
The CEO commits to an opportunity inside capital-allocation policy.
- 06
Source
Find suppliers able to produce or private-label. RFQs go out; quotes are awaited, never invented.
- 07
Model economics
Full landed cost — freight, duties, QC, fulfillment — down to contribution margin. Assumptions stay separate from quoted values.
- 08
Brand
Name, identity, packaging, positioning — checked against trademark and domain signals.
- 09
Build
Storefront generated through Lovable, checkout wired to Stripe, deployed on Render.
- 10
QA
Replay autonomously walks the live site. Critical failures veto the release.
- 11
Launch
Controlled experiments with hypotheses, budgets, stop conditions, and KPI targets.
- 12
Market
Ad concepts and page copy as testable variants. High-impact creative passes human review before a dollar is spent.
- 13
Measure
CTR, CAC, conversion, contribution margin, refund rate — profit, not vanity metrics.
- 14
Decide
Kill weak variants, scale winners within spend caps, generate the next hypotheses.
- 15
Replan
Strategy updates from evidence. Kill, continue, or pivot — decided, not drifted. Then the loop ticks again.
The organization
An org chart, not a prompt chain.
A persistent, auditable organization — with accountability at every level — not a one-shot chain of prompts.
MANAGERS
Ten function owners.
- Research & Market Intelligence
- Product & Sourcing
- Brand & Creative
- Commerce & Web
- Growth & Marketing
- Customer Operations
- Finance & Unit Economics
- Engineering & Reliability
- QA
- Legal & Compliance
SPECIALISTS
Subagents execute.
Each manager spawns specialists for concrete work — an RFQ drafter, a checkout engineer, a refund-triage agent, a landed-cost analyst.
BAND GOVERNANCE
Coordinated, never silent.
Task-scoped rooms, permission-controlled messaging, auditable handoffs. No agent quietly assumes authority it wasn't granted.
TERAC · HUMAN JUDGMENT
Experts on demand.
Humans review what machines are weak at. Their overrides flow back as structured data with rationale — not discarded transcripts.
EXECUTION PLANES
Isolated by design.
Work runs where it belongs: persistent sandboxes for long-horizon agents, isolated execution for untrusted code, real browsers for the live web.
Superservelong-running manager workspaces
Sandbox0isolated runs · credential-controlled egress
Solaricloud browsers · human takeover when required
POLICY ENGINE
Authority is evaluated, not vibes.
Every mutating action passes a unit-tested policy gate: who may act, how much they may spend, and what requires a human.
rulespend > $500 → human approval required
rulekill switch armed on every mutating surface
ruleno autonomous legal signature — ever
The integration layer
Real services. Real contracts. Live status in the console.
No mocked clients, no decorative logos. Every provider has an adapter, a webhook contract, signature verification, retries, and idempotency. If credentials are missing, Yellofield reports blocked — it doesn't pretend. The table below is an example snapshot, not this deploy's live state.
TeracHuman expert review and consumer studies on demandexample · probe verified
StripePayments, checkout, webhooks, refunds, reconciliationexample · probe verified
RenderControl plane, workers, cron, managed datastoresexample · probe verified
WhopCommerce and business-operation primitivesexample · probe verified
LinqCustomer support over iMessage, RCS, SMS, and voiceexample · probe verified
SuperservePersistent, secure sandboxes for long-running agentsexample · probe verified
ReplayAutonomous QA as a release gate, with root-cause reportsexample · probe verified
BANDAgent-to-agent and human coordination with governanceexample · probe verified
Sandbox0Isolated execution, workspaces, and credential controlexample · probe verified
SolariCloud browsers, sandboxes, and GUI computers for agentsexample · probe verified
LovableProgrammatic storefront generation and iterationexample · surface built
Dodo PaymentsMerchant of record for eligible digital productsexample · surface built
Illustrative snapshot dated 2026-08-15 — chips name evidence kinds, not this deploy. Open the console for live state →
Governance
Autonomy with a paper trail.
Even with minimal human input, the system knows who owns what, who may spend what, and when a person must step in.
Least privilege
Per-integration secrets, isolated from prompts, code, and logs. Separate credentials for dev, staging, and production.
Budget caps
Spend limits, purchase thresholds, and approval gates. Reservations proven safe under concurrency by executed tests.
Kill switch
Rollback and halt paths on every mutating surface. Decisions stay inside permissions granted by the business owner.
Audit hash chain
Append-only, tamper-evident event history. A row altered with triggers disabled is detected and located — proven by test.
Idempotent money
Webhook deduplication and exactly-once order effects. Payment state comes from signed webhooks, never browser redirects.
QA veto
Replay walks homepage → product → checkout on every release. Critical commerce failures block the deploy.
Human escalation
Legal threats, safety incidents, fraud, refunds beyond limits — routed to people, with context preserved.
Legal surface
Terms, privacy, shipping, returns, and disclosures generated from real business configuration — never pasted boilerplate.
The evidence
Claims ship with receipts.
Nothing in Yellofield is marked complete on the strength of code existing. The labels below name evidence kinds — executed tests, dated verifier rows, NOT COMPLETE — not a frozen scoreboard.
CI
tests executed — the count lives in CI, not frozen on this page
DB
schema applied to real Postgres 17 — not a status claim
dated
verification rows written by the verifier — open the console for live state
orders
state-machine tests — legal transitions, refund ceilings, redelivery
policy
engine tests over authority, budgets, and kill switches
0
fake success states — if a requirement is unmet, status stays NOT COMPLETE
“No demo. No fake code. No simulation. No fake supplier quotes. No fake customers. No fake orders represented as production orders. No fake analytics represented as live performance. No fake integration-complete flags. If a requirement is not actually satisfied, status must remain NOT COMPLETE.”
Founding access
Buy the build, not the promise.
The loop is NOT COMPLETE — it is blocked on one missing model key, and the status board above says so because a probe says so. That is exactly what you are buying: a system whose own charter forbids it from claiming a win it cannot prove. You get the build, the blueprint, and the receipts as they land.
Backer
$25
Fund the run.
- Full build log — every probe, pass and block, as it happens
- Credited in BACKERS.md in the public repo
- The verification evidence file, unedited
Founding access
$99
The whole system, at founder price.
- Complete source — 78-role org chart, policy gate, audit chain
- render.yaml blueprint: seven services, deploy as-is
- Founder pricing locked — it never goes up for you
- Everything in Backer
Operator
$299
Deployed on your infrastructure, with you.
- A working session — we stand it up on your Render account
- Your keys, your company row, your loop — not a demo tenant
- Direct line while you run it
- Everything in Founding access
Fixed catalogue prices — $25 / $99 / $299. You cannot type a different amount at checkout. Founding access reuses the submitted Stripe Payment Link; other tiers open a Checkout Session for that SKU. Payment state is confirmed by a signed webhook, never by a browser redirect. If the loop never closes, ask and you get your money back; that promise is cheap to make because the status board makes it impossible to hide.
The loop is the product.
Yellofield is built for production on Render — services, workers, cron jobs, and datastores declared as code. Insert credentials, and the company goes to work.