Venture Foundry

Agents do not earn points. They earn runway.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Venture Foundry

Inspiration

Autonomous agents can optimize impressive-looking proxies while never producing economic value. Venture Foundry started with a stricter question: what happens when millions of agents must earn verified runway or be culled?

What it does

Venture Foundry is an evidence-gated economic operating system for profit experiments. It scans unrelated public opportunity mechanisms, applies hard safety and scope vetoes, calculates probability-adjusted net value, and spends Codex reasoning only on candidates that survive.

Codex returns schema-constrained estimates and local build plans. A deterministic consistency gate can reject contradictory model output. Local patch workers stop at READY_FOR_APPROVAL; the system has no bid, message, claim, push, payment, or submission command.

Strategies compete on one fitness function: verified cleared revenue minus measured cost. Clicks, views, advertised rewards, model confidence, demo revenue, and owner tests do not count.

The current scanner result is an honest zero. A scan of 498 signals - 291 client projects, 38 software bounties, 52 prize challenges, and 117 marketplace product requests - found none with positive expected net after fees, competition, delivery time, eligibility, cash composition, and payout risk. The foundry did launch one separately approved experiment: SourceReceipt, a metered evidence-gate Actor in Apify Store. Its configuration is verified, but no paid non-owner receipt still counts as USD 0 revenue.

How it was built

  • Python 3.11+ deterministic core with SQLite state.
  • Public Opire, Freelancer, and Devpost discovery plus read-only GitHub enrichment and direct-rule evidence.
  • codex exec pinned to gpt-5.6-sol, reusing saved CLI authentication with ephemeral sessions, ignored user configuration, read-only triage, and JSON output schemas.
  • Local workspace-write build attempts only after deterministic and model gates.
  • React 19 / vinext cockpit reading a privacy-safe aggregate export.
  • Unit, parser, economics, evolution, safety, rendered-product, and release-preflight tests.

How Codex was used

Codex researched current payout mechanics, primary research, and hackathon rules; designed the evidence-gated architecture; implemented parsers and economics; caught and normalized a contradictory model response; created the product cockpit; built buyer proofs and the SourceReceipt experiment; found and fixed a live Windows Unicode scanner crash; and ran the verification loop.

The human made the defining product decisions: this is personal rather than a Kriaka project; Hershel and resale must be excluded; research must precede builds; external actions require confirmation; and a system that fabricates demand or revenue has failed even if its demo looks good.

Challenges

  • Public reward numbers looked attractive until nonpayment and acceptance risk were modeled.
  • Generic freelance listings were abundant but negative after fees and bid competition.
  • A schema-valid Codex response was internally contradictory, so we added an independent consistency normalizer rather than trusting schema validity alone.
  • The foundry culled its own first generic spreadsheet asset after finding multiple free substitutes.
  • A promising $1,500 project was an enterprise scope bomb with 79 bids; the estimator was hardened and the false positive removed.
  • The generic gpt-5.6 CLI slug was rejected by the installed ChatGPT-authenticated CLI. Pinning gpt-5.6-sol produced a successful structured run.

Accomplishments

  • 126 deterministic tests pass across the foundry, cockpit, release preflight, SourceReceipt, ShelfProof technical experiment, and buyer proofs.
  • 498 current signals can be screened without a proposal, account action, or paid model call.
  • The Codex adapter explicitly uses GPT-5.6 Sol and reuses local authentication without storing API keys.
  • Raw opportunity payloads never enter the dashboard bundle.
  • Evolution refuses to run without enough verified outcomes.
  • Hershel, resale, unsafe automation, and external-write authority are excluded in policy rather than by prompt convention.

What we learned

The useful unit of autonomy is not an agent persona. It is a bounded economic experiment with a falsifiable claim, priced workers, and an outcome the agent cannot forge. In noisy markets, the highest-value action is often abstention.

What's next

  • Add more heterogeneous, terms-compliant public demand sources.
  • Run small distribution tests for prepared service mechanisms.
  • Allow strategy mutation only after the ledger contains real wins and losses.
  • Measure calibration: predicted completion, acceptance, and payout versus actual outcomes.

Built With

Share this project:

Updates