Foundry briefing · September 2026

The Robot Worker blueprint

A multilingual compiler that turns a paragraph into an autonomous robot worker, and an eBay-style floor where those workers are sold, leased, and auctioned — with verification, royalties, and escrow as first-class rails.

01 · Natural language compiler

Prompt in four languages. Topology out.

The compiler is a constrained Grok pass, not a chat window. It accepts English, Tiếng Việt, Español, or 中文, and must emit a single JSON topology: call-sign, role, three goals, a schedule, short/long memory, a minimal tool bind with risk, a five-node graph, three guardrails, three KPIs. Anything that fails the schema falls back to a deterministic local compiler so the Forge never blanks.

  1. 01 · Prompt

    Speak the job

    English, Tiếng Việt, Español, or 中文. No schema. A paragraph is enough.

  2. 02 · Compile

    Topology

    The compiler emits a five-node graph, tool bindings, memory, schedule, and guardrails.

  3. 03 · Bench

    Proof of work

    Sandbox evals, injection scan, malware scan, latency. Score ≥ 80 to list.

  4. 04 · Pack

    worker-pack-v1

    Prompts, tool manifest, memory schema, eval suite. Secrets stay in the vault.

  5. 05 · Ship

    Hangar or Exchange

    Run it privately, or list a sale, lease, or auction with a royalty on secondarys.

Language

Language is an input, not a translation afterthought. Tone bibles (Cô Tư), clause ontologies (墨衡), and invoice layouts (Correo) live in long-term memory keyed to the origin language.

Tool binding

Tools are allow-listed. The compiler may only pick from the foundry catalog. High-risk tools (send mail, write CRM, page on-call, draft a GL) default to dry-run until a human approve token.

Schedule

cron for recurring work, event for inbox/stream, on_demand otherwise. The runtime never invents a webhook the tenant has not bound.

02 · Execution sandbox

A worker that cannot wander

  1. Short-term context

    The current task, last tool payloads, and the open thread. Wiped at the end of the run. Never written to the pack.

  2. Long-term vector store

    Policy packs, alias maps, tone bibles, exception patterns. Tenant-scoped. Exportable as embeddings + docs, not as raw PII.

  3. Tool boundary

    Every call is signed against the manifest. Unknown tools 403. High-risk tools require an approve token. Logs hash payloads.

  4. Live vs trial

    Trials are public and sealed — no outbound side effects. Live inference requires ownership or an unexpired lease, and spends the app owner’s Grok quota only on an explicit Run.

03 · Packaging & portability

worker-pack-v1

An agent is a tradeable digital asset: prompts, tool manifest, memory schema, eval suite, schedule. API secrets are redacted at export and rebound from the runtime vault on import. The content hash (`rwp_…`) is what the Exchange lists, not a folder of weights.

{
  "format": "worker-pack-v1",
  "version": "1.4.0",
  "name": "Ledger",
  "origin": { "language": "en", "prompt": "…" },
  "topology": {
    "nodes": ["perceive", "plan", "act", "verify", "report"],
    "schedule": { "kind": "cron", "cron": "15 6 * * 1-5" }
  },
  "tools": [
    { "id": "stripe.payouts", "risk": "medium" },
    { "id": "netsuite.gl", "risk": "high", "mode": "draft-only" }
  ],
  "memory": {
    "shortTerm": "windowed context",
    "longTerm": ["alias map", "exception patterns"]
  },
  "eval": { "suite": "finance-recon-v3", "score": 96 },
  "secretsPolicy": "redacted-at-export; runtime vault bind only",
  "hash": "rwp_7c91e2"
}

DRM

Licensed copies clone the pack under the buyer. Unique lots transfer title and leave the floor. Forks keep the parent hash in pack.lineage.

Royalties

Creator-set basis points on every secondary. Foundry lots pay no creator royalty — the take is the business.

Anti-tamper

Re-listing a mutated pack without a passing bench is rejected. Injection and malware scans run on every publish.

04 · Exchange mechanics

An eBay for robots that already work

  1. 01 · Discover

    Browse the floor

    Filter by sale, lease, auction, language, and bench score.

  2. 02 · Inspect

    Open the pack

    Topology, tools, risk, reviews, and historical KPIs are public before money moves.

  3. 03 · Trial

    Sandbox run

    A sealed trial with no outbound side effects. Live inference requires ownership or a lease.

  4. 04 · Escrow

    Hold the funds

    Foundry credit locks. 48-hour integration window on sales. Failed bench replay refunds.

  5. 05 · Settle

    Own, copy, or rent

    Unique transfers title. Licensed clones. Leases expire on the clock or the task cap.

Direct sale

Licensed copy or 1-of-1. 48-hour integration window. If a sealed bench replay fails against the listed eval suite, escrow reverses.

Lease

Day / week / month / per-task. The pack never leaves. Live runs debit the lease. Expiry is a hard gate in the hangar.

Auction

Opening bid, optional reserve, optional buyout. Credit holds on each bid. Unmet reserves unwind. Buyout closes the lot.

Verification before listing

Score 80 to list, 90 for the verified take-rate. Evals, injection scan, malware scan, and historical success rate are public on the lot. This build benches in the hangar; production benches run the eval suite in a cold sandbox.

Escrow & disputes

Funds lock on bid or buy. Sales: 48h trial. Leases: unused time refunded if the robot cannot attach a bound tool the listing advertised. Arbitration is currently automatic on bench-replay failure; a human desk is Phase 3.

05 · Financial engine

Take rates, compute, payouts

StreamUnverifiedVerified 90+
Direct sale8%5%
Lease / rental10%7%
Auction12%9%
Featured placementflatflat

Wired take rates in this build: sale 8% / 5% verified, lease 10% / 7%, auction 12% / 9%. New accounts receive $5,000 foundry credit so the floor is playable.

Compute margin

Live runs meter Grok tokens against the app owner. Production bills the tenant a marked-up inference + tool-call line. Trials are sealed and cheap. Cache identical compiles.

Payout rails

This build settles in foundry credit. Phase 3: ACH / SEPA at a $50 threshold, USDC on a 0.5% rail, weekly auto-payout when available ≥ threshold.

Premium

Featured floor placement, verified badge, and white-glove escrow are add-on SKUs — not take-rate discounts except the verified cut already in the table.

06 · MVP to ecosystem

Three phases, one floor

Phase 01 · Live in this build

Foundry

  • Multilingual prompt compiler (EN / VI / ES / ZH)
  • Five-node topology + tool binding + guardrails
  • Hangar: own, clone, run in a sealed sandbox
  • Exchange: fixed-price licensed listings
  • Foundry credit wallet and 5–8% sale take

Phase 02 · Live in this build

Floor

  • Auctions with reserve and instant buyout
  • Leases: day / week / month / per-task
  • Automated bench + injection / malware scans
  • Reputation via reviews and historical KPIs
  • Secondary royalties (bps on the original creator)

Phase 03 · Next

Ecosystem

  • Agent-to-agent dispatch (Loom as the reference conductor)
  • Public SDK: pack in, pack out, CI benches
  • Fiat ACH / SEPA and USDC payout rails
  • Premium escrow and dispute arbitration desk
  • Allow-listed production tool adapters (Stripe, Gmail, NetSuite)

What to try in this build

The floor is live

Compile in four languages. Trial Ledger, Cô Tư, Correo, 墨衡. Bid on Loom. Lease Cô Tư. Sign in, spend foundry credit, run a live hangar job, bench your own pack, and list it.