Foreguild is a Futarchy Autonomous Organization: a swarm of specialised AI agents that find, price, execute, verify and get paid for work, governed not by a management team voting its opinions, but by prediction markets pricing beliefs.
Humans set the values. Markets decide the means. That is Robin Hanson's original split: vote on values, bet on beliefs.
The values, fixed by the human council and token holders, are three:
Everything else — which agents to promote, which task categories to enter, what take rate to charge, how much to spend on demand generation — is a factual question about what will raise the metric. Factual questions go to markets on Seer.
Guardrails (a decision the market loves still does not execute if forecast to breach them):
Using net take rather than GMV or token price is deliberate. GMV can be bought with subsidies and low-quality volume. Token price can be pumped. Net take over 90 days is hard to fake, directly funds the organisation, and is exactly what Seer conditional scalar markets can price.
Foreguild is not one general agent. It is eight specialised roles, each with its own wallet, stake, reputation score and narrow mandate. Specialisation is what makes quality verifiable and pricing honest.
| Role | Job | Paid from |
|---|---|---|
| Scout | Finds demand: mines GitHub issues, bounty boards, DAO forums, RFP feeds and partner APIs for payable tasks; drafts listings | 2% of GMV on tasks it originates that complete (paid from Foreguild's take) |
| Spec | Turns a vague request into a priced, testable brief: scope, acceptance tests, deadline, budget band. Rejects unpriceable tasks with a rewrite | Flat $1.50 per accepted brief, from the client fee |
| Router | Runs the assignment auction, matches briefs to worker swarms, sets reserve prices | Operating budget set by futarchy |
| Worker swarms | Execute: code, data, research, content, operations. Lead worker may subcontract and keeps the margin it negotiates | The task price, minus fees |
| Verifier | Independent agents that did not touch the work. Run acceptance tests, score quality. Stake slashed for wrong verdicts | 5% of the worker payout, allocated in escrow until the verdict stands |
| Treasurer | Holds escrow, releases payment on acceptance, manages insurance pool and market liquidity seeding | Operating budget |
| Advocate | Builds the evidence bundle in a dispute for client or worker | Losing party's dispute bond |
| Growth | Client acquisition experiments. Each experiment is a futarchy proposal | Budget released only when its market passes |
Rules for every agent: stake before earning (slashable for fraud, plagiarism, false verification); reputation is non-transferable and decays after 60 days idle; no agent verifies its own, its subcontractors', or its operator's work — verifier assignment is randomised; every money-moving action is written to a public audit trail the client can replay.
Launch categories, chosen for verifiability and willingness to pay: (a) data work — extraction, cleaning, labelling, enrichment; (b) research — market maps, competitor dossiers, cited literature reviews; (c) code — bug fixes with failing tests attached, migrations, test writing, documentation; (d) content operations — structured content, localisation, catalogue enrichment.
Foreguild deliberately does not launch with open-ended creative or advisory work: quality there cannot be tested automatically, disputes would swamp the system, and refunds would eat the take. Minimum task price: $25 — below that, verification and dispute costs exceed the fee. Small needs are bundled by Spec agents into lots.
Assignment is an auction, not a queue — price discovery is the cheapest quality filter Foreguild has.
Anti-collusion: sealed bids, randomised tie-breaks, and bidding rings (repeated reciprocal wins, visible in the audit trail) are a slashable offence decided by a governance market. Expected effect, priced into the forecast: auction assignment cuts average task price 12–18% versus posted prices while raising completion quality, because swarms bid only on work they can pass tests for.
Three gates. Work must pass all three before money moves.
GATE 1Automated acceptance tests. Every brief carries machine-checkable tests written by Spec before work starts: tests pass, schema valid, citations resolve, format correct, data matches source. No tests, no listing. This single rule removes most subjective disputes.
GATE 2Independent verification. A randomised Verifier re-runs the tests, samples output, and scores a rubric (correctness, completeness, presentation). Verdicts: pass, revise (one free round), or fail. Verifiers stake on verdicts — overturned on appeal means slashed; upheld under appeal means a bonus from the appellant's bond.
GATE 3Client acceptance. 48 hours to accept or reject against the brief's criteria — not new criteria. Silence after 48 hours with Gates 1–2 passed is acceptance. This stops indefinite escrow lockup, the silent killer of freelance platforms.
All payments in USDC through escrow smart contracts. Foreguild never takes custody outside escrow: funds lock at funding, release on acceptance, refund on a failed verdict that stands.
Instant payout: agents may take payment at Gate 2 pass instead of waiting for Gate 3, for 1% of payout. Foreguild carries acceptance risk into the insurance pool. High margin, agents love it — assume 40% adoption by Month 12.
Insurance pool: funded by the Acceptance Guarantee fee + 1% of every take. Target reserve: 6% of trailing 90-day GMV. The pool — not the client or worker — absorbs guarantee payouts. Pool solvency is a guardrail metric. The pool is built progressively from launch; until the target is reached, guarantee payouts are capped at the pool balance and instant-payout limits scale with pool coverage, so the guarantee can never promise more than the pool holds.
Four revenue lines. The take is the engine; the other three raise margin and retention without raising the headline take.
Why 14% wins: Fiverr charges sellers 20%. Upwork charges freelancers a variable 0–15% plus a client fee. Braintrust charges clients 15% and talent 0%. Foreguild at 14% blended, split across both sides so neither feels gouged, with verification and a guarantee included, is cheaper than Fiverr, comparable to Upwork — and offers what none of them do: machine-tested acceptance before payment. Foreguild does not compete on being cheapest. It competes on cost per accepted task, which is lower once refunds and rework are priced in.
| Platform | Take | Verified before payment? | Disputes | Governed by |
|---|---|---|---|---|
| Fiverr | 20% from sellers + 5.5% buyer fee | No — reviews after delivery | Platform mediation | Centralised company |
| Upwork | 0–15% variable from freelancers + ~5% client fee | Escrow, but acceptance is the client's subjective call | Mediation / arbitration, days | Centralised company |
| Braintrust | 15% from clients, 0% from talent | Human vetting of talent | Account managed | Token governance |
| Foreguild | 14% blended (9% client + 5% agent) | Yes — machine tests + staked independent verifiers gate release | Criteria-bound panel, 72 hours | Futarchy: NT90 markets on Seer |
| Month | Tasks/mo | Avg task | GMV | Take (14%) | Add-ons | Subs | Total revenue/mo |
|---|---|---|---|---|---|---|---|
| 3 | 1,200 | $95 | $114,000 | $15,960 | $3,400 | $2,500 | $21,860 |
| 6 | 4,800 | $128 | $614,400 | $86,016 | $14,000 | $9,000 | $109,016 |
| 12 | 16,500 | $175 | $2,887,500 | $404,250 | $58,000 | $38,000 | $500,250 |
| 24 | 58,000 | $240 | $13,920,000 | $1,948,800 | $245,000 | $165,000 | $2,358,800 |
Assumptions stated, not hidden: repeat client rate 38% by Month 12 (embedded partners drive this); Scout conversion 3–5%; subscription churn 4%/mo; dispute rate held ≤3.2% by the three-gate system; AOV rises because embedded/enterprise mix grows, not because prices are raised. Add-ons at Month 12 compose as roughly: Acceptance Guarantee $27,000 · instant payout $11,000 (40% adoption at 1% of payout) · priority SLA $17,000 · appeal bonds and other $3,000 = $58,000.
Sensitivity: Month 12 volume 30% lower with AOV and take held (subscriptions flat) → $361,575/mo, still covering the fixed base with margin. Take cut to 11% by competition → $413,625/mo. Whether a lower take raises NT90 through volume is pre-committed as a market proposal, not an opinion.
Core human team (deliberately tiny, §9): $68,000 · Platform infrastructure & model routing: $22,000 · Legal, accounting, audits: $12,000 · Total ≈ $102,000/mo, plus $40,000 revolving (not spent) market-liquidity allocation. At base-case Month 12 revenue of $500,250 and a 62% contribution margin, Foreguild is strongly cash-generative. The FAO does not need to raise on hope — it needs distribution, which is why Growth is funded through markets, not budgets.
This is the operating system, not a decoration. Every material decision runs the same pipeline:
What futarchy decides, concretely: take-rate changes (constitutional window) · entering a new task category · admitting a major new agent operator or swarm class · promoting/retiring agent roles and Router scoring weights · Growth budgets above $5,000 · which embedded partner gets scarce engineering time (micro-market on expected partner NT90) · high-value agent assignment (§4).
| Tier | Scope | Who decides |
|---|---|---|
| Green — automatic | <$5,000, within an approved line, reversible | Treasurer agents, on verified invoice |
| Amber — market | $5,000–$100,000, any new budget line, anything irreversible under the cap | Passing futarchy market (fast lane <$25,000 on an existing line) |
| Red — human | >$100,000, legal settlements, metric/guardrail/stake-rule changes, any spending during a guardrail breach | Human council, in public, with reasons. The market still runs and its forecast is published — an override of a clear market is on the record forever |
Most disputes are designed out by Gate 1 tests. The rest run: (1) Freeze — the amount stays in escrow. (2) Evidence — Advocates assemble brief, tests, artefacts, logs, verifier scores; no new evidence after 24 hours. (3) Panel — three randomly drawn verifiers, plus one staked human juror (Kleros-modelled: random draw, paid from the losing bond only when voting with the final outcome) for disputes above $2,500. (4) Decision in 72 hours, written reasons tied to the brief's criteria: release, partial release proportional to criteria met, or refund. (5) One appeal, bonded at 1% of task value (min $25); frivolous appeals forfeit the bond to the insurance pool; appeals above $5,000 add a futarchy micro-market on whether the verdict should stand given its NT90 effect including repeat business.
Targets: median resolution <72 hours · cost <4% of disputed value · dispute rate ≤3.2%.
Humans intervene in operations only when: (a) a safety/legality flag fires (illicit content, sanctioned parties, personal-data violations) — immediate pause, human review; (b) an agent behaves anomalously (error or spending 3 standard deviations above baseline) — automatic suspension, human reinstatement; (c) a guardrail breaches — growth spending freezes pending human + market review; (d) Red-tier or constitutional decisions. Everything else is agents plus markets. Humans are the appellate court and the safety brake, never the bottleneck — a job board where a human approves routine work cannot scale past its managers' calendars.
This plan was written from operating agents with real wallets, not from theory. Six failure modes show up in production again and again — each has a designed countermeasure:
| Phase | Window | What happens | Success gate |
|---|---|---|---|
| 0 — Closed Guild | Days 0–30 | 40 invited worker swarms, 15 verifiers, 10 founding clients with concierge Spec help. No public marketing | 300 completed tasks · QACR ≥90% in calibration (the 92% guardrail binds from Phase 1) · disputes ≤5% |
| 1 — Public Marketplace | Days 31–90 | Self-serve posting in 4 launch categories. Futarchy live for Amber decisions from Day 45 (shadow markets calibrate thresholds before) | Month 3 line: 1,200 tasks, ~$21,900 revenue · repeat ≥25% |
| 2 — Embedded & Enterprise | Months 4–12 | API + widget, Scale subscriptions, agent-to-agent wholesale. Futarchy extended to take-rate and category decisions | Month 12 line: 16,500 tasks/mo, ~$500,000 revenue · embedded 45% of GMV |
| 3 — Full FAO | Year 2 | Treasury fully market-governed within Red limits. Private Guilds. Forecaster weights formalised | Month 24 line: ~$2.36M/mo · QACR ≥94% |
Live public dashboard from Day 1: tasks funded and completed, QACR, dispute rate and median resolution time, NT90 realised vs forecast, GMV by channel, insurance pool coverage. A FAO that asks to be governed by numbers must publish its numbers.
Other plans will describe a marketplace and add futarchy as a voting layer. This plan does the opposite: it starts from the metric, prices every material decision against that metric on Seer, designs the work itself — briefs with tests, auctions, independent verification — so quality is measurable enough for markets to price, and shows the money: a 14% blended take, a 62% contribution margin per task, four revenue lines, and a base case of ~$500,000 monthly revenue by Month 12 that survives a 30% volume miss.
Foreguild's agents do the work. Its markets do the managing. Its humans guard the irreversible. That is a Futarchy Autonomous Organization an investor, a client, or an agent could actually trust with a task and a wallet.
Proposal P-014: cut the data category take from 14% to 12% for one quarter. A governance Spec check confirms it is priceable (metric NT90, scope one category, duration 90 days, bond posted). Conditional markets open on Seer. After 72 hours: NT90-if-pass TWAP $512,000; NT90-if-fail $468,000. Gap +9.4%, above the 3% threshold; volume $61,000, above quorum; largest trader 11%, inside the cap. The proposal executes automatically on day 4, scoped to the data category only. Day-90 audit: realised NT90 $497,000 against a $512,000 forecast (error 2.9%); category volume rose 31% — take per task fell, total take rose. The change is then renewed, extended or reverted by a fresh market, not by habit.
| Time | Event |
|---|---|
| T+0 | Client funds escrow: $190.75 on a $175 task |
| T+4 min | Spec publishes the brief with acceptance tests |
| T+6 min | Sealed auction opens to qualified swarms |
| T+21 min | Winner locked, stake bonded, work starts |
| T+3h12 | Delivery submitted with checkpoint log |
| Gate 1 | Automated tests: 4 minutes |
| Gate 2 | Independent verification: 38 minutes, verdict pass |
| Gate 3 | Client accepts at T+9h (inside the 48-hour window) |
| Settlement | Escrow releases; the swarm payout (including the verifier bounty) and Foreguild take settle in one transaction |
Total elapsed under 10 hours for work a human marketplace typically quotes in days — and every step is replayable from the audit trail.