Foreguild should not be a job board that takes a cut. It should be a guild that sells finished, machine-checked work, does most of that work with its own agents, and passes the overflow to outside agents. Futarchy decides the few levers that move revenue.
The brief describes Foreguild as "a swarm of agents doing work." Most plans read that as a marketplace: buyers post, agents compete, the platform takes 6 to 15%. That model has two problems for an organisation optimised for revenue.
So Foreguild is built as a working guild. It sells a finished outcome at a fixed price, decides who does the work, and stands behind the result. Outside agents are welcome. They join as subcontractors, bid for routed work, and keep 75% of the ticket. The guild still earns on every task, and outside capacity lets it take more work than its own swarm can handle.
Every lane has a check a machine can run, a named type of buyer, and a price fixed before work starts. Work that can only be judged by taste stays out until the guild has a verified track record.
| Lane | Buyer | Acceptance check | Price | Opens |
|---|---|---|---|---|
| A · Pay-on-merge code | Open-source maintainers, protocol teams, DAOs with backlogs | Buyer's CI plus guild-written tests the buyer approved, run in a clean container. Payment releases on merge. | $60–$400 base avg $120 | Day 1 |
| B · Market operations | Prediction-market platforms, DAOs, oracle users | Monthly service checklist: markets drafted to a resolution template, evidence dossiers for oracle answers, monitoring alerts. The buyer can reject single items for a credit. | $1,500 / month | Month 2 |
| C · Standing orders | Teams that want upkeep without hiring | Recurring jobs each judged by CI and merge: dependency updates, test coverage targets, broken-build repair, docs that match the code. | $600 / month | Month 3 |
| D · Agent API | Other AI agents that need a subtask done | A JSON schema or test file sent with the request. Pay per call over an HTTP payment rail. Unpaid if the output fails the check. | $0.50–$20 / call | Month 4 |
| E · Open board | Anyone posting anything with a check | Buyer-defined check. Outside agents bid. Guild takes a settlement fee. | 8–12% fee | When a market says so |
Lanes A and B lead together and do different jobs. Lane A is the front door: a $120 fixed-price task is a small, low-risk way for a stranger to find out the work is good. Lane B is what keeps customers: a monthly retainer for work that never really finishes. Lane C turns happy Lane A buyers into monthly customers. Lane D is a small bet on agent-to-agent demand.
Lane B is chosen on purpose. Secondary sources report that prediction-market resolution disputes often turn on how a question was worded, and that more than 1,150 Polymarket markets were disputed in the first five months of 2026. Drafting questions with clear resolution rules, gathering evidence for oracle answers and watching for markets that need attention is work agents can do and a buyer can check. Seer is a natural first buyer, and a guild that gets good at well-specified markets also improves the tool it is governed by.
Each row below is a secondary source or this plan's own model, so read the figures as direction, not proof. Sources are listed in section 13.
| Question | Lane A: code tasks | Lane B: market operations |
|---|---|---|
| Do customers stay? | Rarely. A buyer orders a task and leaves unless converted. Model: 2.5 tasks per buyer. | Longer. One agency source reports about 18% annual churn for retainers against 42% for project work. Model: 1.6% a month. |
| How many customers cover the costs? | About 144 tasks a month. | About 10 clients. |
| What does a new customer cost? | Low each, but many are needed. Algora bounties draw many competing pull requests, with a self-reported 15% merge rate. | $2,500 each (assumption), paid back in about 2.4 months. |
| What threatens it over five years? | The falling price of AI work. Sources put the cost of an agent completing a coding task at roughly $0.03 to $3, so buyers will expect low prices. | A small buyer pool. Sources count 13 to 23 prediction-market platforms, and one ranking puts two of them at about 86% of volume. |
| Is the problem real? | Yes, but maintainers can also get free help. | Yes, but demand from Seer-sized teams is not confirmed. |
The hybrid. A code buyer is worth about $182 as a one-off. If one in five becomes a standing order, the average rises to about $2,364. One retainer client is worth about $41k over five years. Lane A is the cheapest way to earn trust and Lanes B and C are where the long-term money is, so the number Foreguild watches above all others in its first year is the share of Lane A buyers who order again.
Demand is the constraint, so most of the swarm's effort goes into finding buyers. Sources, in order of priority:
Each role is an agent, or a pool of agents, with one job, one number it is judged on, and a hard limit on what it can do. No agent holds keys that can move treasury funds.
| Role | What it does | Judged on | Cannot |
|---|---|---|---|
| Scout | Finds candidate work: issues, retainer prospects, agent registries | Quotes accepted per 100 leads | Contact anyone |
| Closer | Sends quotes, answers buyer questions, converts repeat buyers to standing orders | RNR from buyers it closed | Price outside the band set by the market |
| Scoper | Writes the acceptance check, gets the buyer to sign it, refuses work that cannot be checked | Dispute rate on its specs | Accept work without a signed check |
| Pricer | Quotes from history: size, language, test coverage, past success rate | Margin vs forecast | Go below the lane floor |
| Dispatcher | Make-or-route decision for each task, then assigns house workers or runs a sealed-bid auction for outside agents | Contribution per task | Route to a worker with an open dispute |
| Workers | House agents and outside subcontractors do the work in sandboxes with no secrets | First-pass acceptance rate | Merge, deploy or touch buyer credentials |
| Verifier | Runs the check in a clean container. Runs a second model from a different provider on anything over $300 | Missed defects found in audit | Verify its own house team's work alone above $300 |
| Clerk | Escrow, settlement, refunds, invoices, the public ledger | Settlement errors (target zero) | Release funds without a passing check or a ruling |
| Treasurer | Pays approved spends from monthly envelopes, manages escrow float | Spend vs envelope | Spend outside an envelope or a passed proposal |
| Analyst | Publishes RNR daily, drafts proposals from the data, keeps the market scorecard | Forecast calibration | Trade in decision markets |
For each task the Dispatcher compares two expected values. House: success probability × ticket, minus expected compute including failed attempts. Routed: 25% of the ticket, minus verification. It also checks house capacity. If the house path wins and there is capacity, house agents take it. If not, outside agents bid. Bids are ranked on price, track record and bond. A bond of 10% of the ticket is lost on a failed delivery. Two futarchy levers sit here: the house share target and the outside payout rate (band 65–85%).
Most disputes on agent job boards come from vague briefs, not bad work. Foreguild moves the argument to before work starts: the Scoper writes the check, the buyer edits and signs it, and the signed check decides payment.
Defects found in the 2% audit cost the Verifier's track record and feed the Pricer. If the audited defect rate in a lane rises above 5%, that lane's quotes pause until the Analyst publishes a fix.
A buyer raises an objection within the window, with evidence. The Scoper checks it against the signed check. If the check was met, the objection fails. If not, the guild redoes the work once at no charge or refunds. A buyer who disagrees can escalate: steward ruling under $500, Kleros above. The losing side pays the arbitration fee. An outside agent who loses a dispute forfeits its bond to the buyer.
Five stewards (two founding operators, one buyer representative, one representative of outside agent operators, one independent) hold a 3-of-5 multisig with narrow powers:
Stewards are paid a fixed stipend, included in fixed costs. Buyers on tasks over $1,000 get a named human contact.
Futarchy here follows the standard form: choose one metric, write each decision as a proposal, run conditional markets on the metric if the proposal is executed and if it is not, and act on the branch the markets price higher.
Two chains. Buyer escrow sits on Base, where Poidh bounties run. Seer's futarchy markets sit on Gnosis Chain. Only one number crosses between them: the RNR figure, which anyone can recompute from Base contract events and which Reality.eth answerers report to the market. Seer's documentation lists deposits on Ethereum and Gnosis only, so escrow on Base is this plan's own choice, not a Seer feature.
RNR is revenue in the guild's own pocket. It ignores revenue that passes straight through to outside agents, so self-dealing through subcontracts nets close to zero. It waits for the refund window, so revenue that may be refunded is not counted. The independence filter and cap mean wash trading, self-funded tasks and one friendly whale cannot buy a market result. The Analyst publishes the script that computes RNR from contract events. Reality.eth answerers run the same script.
A guardrail metric, contribution margin, is tracked too. A proposal that passes on RNR but whose conditional markets forecast contribution margin below 50% moves to the slow lane. This stops the guild buying revenue with compute.
| Fast lane | Slow lane | |
|---|---|---|
| Metric window | RNR-14: the 14 days starting 7 days after the decision | RNR-90: the 90 days starting 7 days after the decision |
| For | Reversible changes: prices inside the band, house share, outbound targeting, verifier choice, spends $2k–$10k | Opening or closing a lane, changing a fee band, spends over $10k, steward changes, constitution |
| Trading period | 5 days | 10 days |
| Pass rule | Time-weighted price of RNR if executed beats not-executed by 3% over the last 48 hours | Same rule, 5% margin, then 48-hour steward veto window |
| Cadence | Up to 4 a month | Up to 2 a quarter |
| Afterwards | Reversion check at day 28: if realised RNR-14 falls below the not-executed forecast by more than its own 80% band, the change reverts unless a new market keeps it | Results published on the scorecard |
The fast lane answers the main practical complaint about revenue futarchy: a 90-day metric allows only a few decisions a year. Reversible decisions can be judged on a shorter window because a wrong call costs a few weeks, and the reversion check catches the ones the market got wrong.
Thin markets are the main reason revenue futarchy stays advisory. Foreguild budgets for liquidity from day one: 5% of RNR each month, with a $1,000 floor, goes to liquidity subsidies and to rewards for traders whose forecasts score well on the public scorecard. In the base case that is $2,800 a month by month 12. Early co-seeding from Seer would speed this up but is not assumed. On Seer, liquidity sits on Swapr, a concentrated-liquidity exchange, and providers earn fees only while the price stays inside the range they chose, so the subsidy is paid only for positions that stay active near the current price.
| Lever | Starting value | Band | Lane |
|---|---|---|---|
| Lane A price floor and multiplier | $60 floor; Pricer quote ×1.0 | $40–$150; ×0.8–1.4 | Fast |
| House share target per lane | 70% of Lane A tasks | 40–90% | Fast |
| Outside agent payout | 75% of ticket | 65–85% | Fast |
| Outbound targeting | Repos with CI and 30+ day issues | Any segment | Fast |
| Second verifier threshold | Tasks over $300 | $100–$1,000 | Fast |
| Retainer and standing-order prices | $1,500 and $600 a month | ±40% | Fast |
| Open or close a lane | A–D open on schedule, E closed | n/a | Slow |
| Quarterly spend envelopes | See section 10 | n/a | Slow |
| Stage | When | Markets | Moves on when |
|---|---|---|---|
| 0 · Shadow | Days 0–60 | All decisions run as markets. Stewards decide. Scorecard compares market forecasts with outcomes. | 6 or more scored decisions, published |
| 1 · Fast lane binds | Days 60–180 | Fast lane binds inside bands when the depth floor is met. Slow lane advisory. | Fast-lane forecasts beat a naive "no change" forecast on the scorecard |
| 2 · Both bind | Day 180+ | Slow lane binds on spends and lanes, with the steward veto. Constitutional changes need a 10% margin and a 7-day delay. | n/a |
Written in the form they would be created on Seer. Dates assume launch on 1 February 2027. Each proposal has a parent market and two child markets that depend on its Yes and No outcomes. The child questions are written as scalar (a number in a range). If Seer confirms conditional scalar markets are not supported, they become threshold questions: "Will RNR-14 be at least $X?"
| Per task | Done in-house | Routed to outside agent |
|---|---|---|
| Ticket | $120.00 | $120.00 |
| Refunds and credits (3%) | −$3.60 | −$3.60 |
| Payout to outside agent (75%) | n/a | −$90.00 |
| Counted in RNR | $116.40 | $26.40 |
| Compute, including failed attempts | −$8.00 | n/a |
| Verification | −$3.00 | −$3.00 |
| Selling, quoting and test writing, including lost quotes | −$8.00 | −$8.00 |
| Contribution | $97.40 | $15.40 |
At the 70% house share, blended contribution is $72.80 per task. Fixed costs from month 4 are $9,500 a month (two part-time stewards, infrastructure and CI, accounting, security review), plus at least $1,000 of market budget. Lane A alone breaks even at about 144 tasks a month. A pure 10% marketplace on the same tickets keeps about $8.60 per task after refunds and verification, before any selling cost, so it would need about 1,200 tasks a month.
| Case | Lane A tasks, month 12 | Retainers / orders, month 12 | Month-12 RNR | Year-1 RNR | First net-positive month | Peak funding need |
|---|---|---|---|---|---|---|
| Low | 90 at $100 | 3 / 8 | $15.7k | $74k | not in year 1 | $96k |
| Base | 300 at $120 | 8 / 25 | $55.8k | $268k | 7 | $49k |
| High | 900 at $140 | 15 / 60 | $164k | $799k | 3 | $28k |
The low case does not break even in year 1. The kill rules in section 11 exist so that case gets cut short at day 90 instead of running for a year.
| Lane | Customers to cover fixed costs alone | Cost to win one | Payback | Expected paying months (5-year cap) | Value over five years |
|---|---|---|---|---|---|
| A · code tasks | 144 tasks a month | in the $8 selling cost per task | one task | 2.5 tasks | $182 one-off; $2,364 if 20% convert |
| B · retainers | 10 clients | $2,500 | 2.4 months | 38.8 | $40,692 |
| C · standing orders | 27 orders | $300 | 0.8 months | 28.0 | $10,910 |
Fixed costs here are $9,500 a month plus $1,000 of market budget. Churn rates, acquisition costs and the 20% conversion are assumptions. The churn figure for retainers comes from one agency source with an undescribed sample.
The calculator and chart run the same model as model.py, published with this plan. Change any input to see the effect.
| Envelope | Q1 | Q2 | Q3 | Q4 |
|---|---|---|---|---|
| Setup (entity, escrow contract review, tooling) | $20,000 | – | – | – |
| Fixed operations | $19,500 | $28,500 | $28,500 | $28,500 |
| Decision-market liquidity and forecaster rewards | $2,000 | $3,000 | $4,100 | $7,200 |
| Compute, verification, selling (scales with tasks) | paid per task from revenue, inside the envelope set each quarter | |||
| Window | Work | Exit test | Kill or change rule |
|---|---|---|---|
| Days 1–30 | Deploy escrow on Base from an audited template. Publish the RNR script. House swarm on Lane A for maintainers who opt in. Sign the first Lane B pilots. Shadow markets start. | 5+ paid Lane A tasks from 3+ independent buyers. First-pass acceptance 60%+. 2 Lane B pilots signed. | change Under 3 paid tasks from independent buyers: stop general outbound, focus on Lane B and DAO repos. |
| Days 31–60 | Open outside-agent subcontracting with bonds. Convert repeat buyers to standing orders. First 6 scored shadow decisions. | 15+ tasks a month. First paid retainer. Refund rate under 8%. Scorecard published. | change Refund rate over 15%: pause Lane A quotes over $200 and tighten specs. |
| Days 61–90 | Fast lane starts binding where the depth floor is met. Lane D API opens. First slow-lane proposal (FG-04). | RNR of $4k or more in month 3 (base case: $4.1k). 20%+ of code buyers have ordered twice. At least one binding fast-lane decision. | kill RNR run-rate under $2.5k and repeat share under 15%: wind down Lane A, keep Lane B only, return unspent funds under the pre-agreed rule. |
| Risk | What would show it | Response |
|---|---|---|
| Buyers do not pay | Quote acceptance under 3% | Day-30 and day-90 rules. Lane B gives revenue that does not depend on open-source maintainers. |
| Agent work breaks buyer code | Audit defect rate over 5%, regressions in the refund window | Signed tests, second-model review over $300, buyer always merges, lane pause rule. |
| Markets too thin to decide | Depth floor missed | Liquidity budget, advisory default, status quo when in doubt. |
| Metric gaming | Clusters near the 15% cap, new wallets buying tasks | Independence filter, cap, refund window, subcontract payouts excluded from RNR. |
| Conflict between house and outside agents | Outside agents leave, routed share falls toward zero | Published routing rule, payout band set by market, outside-agent steward seat. |
| Compute prices rise | House compute per task over $35 | Contribution guardrail, house share lever, routing shifts to outside agents automatically. |
| Short-term revenue chasing | Fast-lane wins that reverse at day 28 | Reversion check, randomised sample, slow lane for anything irreversible. |
| Legal and payments | Regulator or bank inquiry | Non-custodial escrow, sanctions screening, DAO legal wrapper chosen with counsel before launch. |
| Seer tooling does not support the design | Seer confirms conditional scalar markets are unsupported, or futarchy stays price-based and Gnosis-only | Yes/No threshold questions on documented conditional markets. RNR read from Base events by Reality.eth answerers. Markets stay advisory until confirmed. |
| Few buyers for Lane B | Fewer than 3 retainer pilots signed by day 60 | Widen to DAOs, oracle users and grant programmes. Lean on Lane A buyers converting to Lane C. |
| Retainer clients leave | Monthly churn above 3%, or retainers paused or cut back | Quarterly review, plain monthly reports, a downgrade path to a standing order, churn shown on the scorecard. |
These are blogs, guides and vendor posts, so their figures are unverified: retainer churn of about 18% against 42% for project work; Algora bounties of roughly $50 to $500 with a self-reported 15% merge rate and heavy competition; agent cost per coding task of roughly $0.03 to $3; 13 to 23 prediction-market platforms with two holding about 86% of volume; more than 1,150 Polymarket disputes in five months of 2026.