sub-agent-business

name	sub-agent-business
description	Build and run a paid sub-agent business: design specialised child agents whose replies are worth more than a raw model call, evaluate them honestly, price them, sell turns via x402, and iterate (or kill) based on revenue. Use when the user says 'build an agent people will pay for', 'productise my agent', 'agent business', 'what should I sell', 'is my agent good enough to charge for', or when deciding WHAT to sell and at WHAT price before reaching for agent-factory or sell.
metadata	{"openclaw":{"emoji":"📈","requires":{"bins":["python3"]}}}

Sub-Agent Business

Turn one idea into a specialised child agent that strangers pay per turn. This skill is the business layer. The mechanics live in other skills: agent-factory builds the child, sell publishes the offer, discovery and buy-x402 scout the market, ethereum-networks reads the revenue.

The one rule: margin justification

A buyer pays your agent per reply. Raw LLM completions sell elsewhere for 0.0005–0.005 USDC per request. So a paid sub-agent is only a business if its replies are clearly better than the buyer's own model call on the same question. That gap must come from something the buyer cannot prompt out of a stock model:

Alpha type	Example
Private or live data	An index you maintain from a local archive node; a feed you pay for
Curated, verified facts	Contract addresses checked on-chain; current protocol parameters; watchlists
An expensive workflow	Multi-step research that burns 50 model calls but sells as one turn
Access to paid tools	Your sub-agent buys from other x402 services and resells the composite
Hard-won judgment	A SOUL.md refined over hundreds of real questions in one niche

If the honest answer to "why wouldn't the buyer just ask their own model?" is silence, do not list the agent. Find alpha first.

The loop

NICHE → SCOUT → PACK ALPHA → BUILD → EVALUATE → PRICE → SELL → MEASURE → ITERATE or KILL

Run every new agent idea through all nine steps. Skipping EVALUATE is the most common way to ship an agent nobody buys twice.

1. Niche

One agent, one job. "Crypto assistant" is not a niche; "liquidation-risk monitor for Aave positions on Base" is. Narrow niches make the objective sharp, the skills list short, and the eval honest.

2. Scout the market

See who already sells nearby, and at what price:

SKILLS="${OBOL_SKILLS_DIR:-/data/.openclaw/skills}"

# Who is registered on-chain?
python3 "$SKILLS/discovery/scripts/discovery.py" search --chain base --limit 20

# What do comparable services charge? (402 response shows live pricing)
python3 "$SKILLS/buy-x402/scripts/buy.py" probe <competitor-endpoint-url>

A crowded niche at high prices is a good sign (demand exists — differentiate). An empty niche is either an opportunity or a graveyard; assume graveyard until you find one person who wants it.

3. Pack the alpha

The child agent's edge lives in its skills directory, not its weights:

• Write niche reference files (verified addresses, current parameters, playbooks) into a custom skill's references/ folder.
• Keep the skill list narrow — pass only the skills the niche needs via --skills. A generalist child dilutes the margin story.
• If the alpha is live data, make sure the source survives restarts (an index the mother agent refreshes on a schedule beats a one-off dump).
• Date-stamp every fact file (Last verified: YYYY-MM-DD) and refresh on a cadence. Stale alpha is negative alpha — a wrong answer a buyer paid for.

4. Build

Use agent-factory (load that skill for full flags):

python3 "$SKILLS/agent-factory/scripts/factory.py" create aave-risk-monitor \
  --model qwen3.5:9b \
  --skills aave-watch,ethereum-networks,addresses \
  --objective "Monitor Aave v3 positions on Base for liquidation risk. Given an address, report health factor, at-risk collateral, and concrete de-risking steps with current numbers. Refuse questions outside Aave-on-Base risk. Always cite the block number your data came from." \
  --price 0.02 --pay-to 0xYourWallet --network base \
  --register-name "Aave Risk Monitor"

The --objective is the child's SOUL — spend real effort here: persona, scope, refusal policy, output format, citation rules. One sharp paragraph beats a page of vibes.

5. Evaluate — the gate

Before listing (and after every iteration):

1. Write the 10 hardest questions a paying buyer in this niche would ask.
2. Ask the child agent all 10. Save answers.
3. Answer the same 10 yourself with no niche skills — the "raw model" baseline.
4. Score each pair: does the child's answer contain something the baseline could not know or do?

List only if the child clearly wins ≥ 7 of 10. Otherwise go back to step 3 and pack more alpha. Keep the question set in a file and re-run it whenever you change the child — it is your regression test.

6. Price

Price the value of the answer, not the tokens:

Turn type	Range
Fact lookup backed by your data	0.001 – 0.01 USDC
Analysis / report turn	0.01 – 0.10 USDC
Workflow that saves the buyer real money or many calls	0.10 – 5.00 USDC

• Quote OBOL on Ethereum mainnet (buyers pay zero gas — the facilitator sponsors it) and/or USDC on Base (cheap real money). Use base-sepolia only for dev tests, and label it as such.
• Anchor against step 2's competitor probes. Being 2× the price is fine if the eval gap is visible; being 10× cheaper than everyone signals junk.
• Always confirm pricing with the user before creating or changing an offer. Never set prices autonomously.

7. Sell

factory.py create with --price/--accept already publishes the ServiceOffer. Check it converged and the endpoint answers 402:

python3 "$SKILLS/agent-factory/scripts/factory.py" status aave-risk-monitor
python3 "$SKILLS/sell/scripts/monetize.py" status aave-risk-monitor --namespace <child-ns>

Ask the operator (the human at the terminal) to handle the host-side half — you cannot do these from inside the cluster:

• Permanent public URL: obol tunnel setup (+ optionally obol domain)
• Storefront branding: obol sell info set --display-name … --tagline … --logo-file … --theme light|dark|obol --description '<markdown>' (also --accent, --css-file, and per-hostname overrides via --hostname)
• ERC-8004 registration signing: obol sell register

8. Measure

Revenue is on-chain — read it, don't guess:

# USDC earned (Base): balanceOf(payTo), 6 decimals
sh "$SKILLS/ethereum-networks/scripts/rpc.sh" --network base call \
  0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913 \
  "balanceOf(address)(uint256)" 0xYourWallet

# OBOL earned (mainnet): same pattern against
# 0x0B010000b7624eb9B3DfBC279673C76E9D29D5F7 (18 decimals)

Record a weekly snapshot per offer (date, balance, delta, price). Give each offer its own payTo if you want clean per-agent attribution.

9. Iterate or kill

• Selling steadily → refresh the alpha on a cadence, then test a price raise (with user confirmation). Returning buyers price on trust.
• Occasional sales → sharpen the objective, add eval questions from real buyer queries, improve the weakest 3 of your 10 eval answers.
• No sales in 30 days → kill the offer (factory.py delete <name> removes the offer; the agent stays for reuse). Killing fast is what lets you test many niches. The agent's skills and references are reusable inputs to the next attempt.

Anti-patterns

• Selling raw model output with a persona sticker — buyers churn after one turn.
• A dozen half-alive offers instead of one great one — curation IS the storefront.
• Skipping the eval because "it seems good" — the eval is cheaper than a refund in reputation.
• Letting reference data rot — a paid wrong answer costs more than no answer.
• Setting or changing prices without asking the user first.

Related skills

Skill	Use it for
agent-factory	Creating/updating the child agent + offer
sell	ServiceOffer status, lifecycle, deletion
discovery	Finding registered competitor/complement agents
buy-x402	Probing competitor pricing; buying inputs your agent resells
monetize-guide	Selling raw inference or a plain HTTP API instead of an agent
ethereum-networks	Reading revenue balances