Everything ComputeLine does, in plain terms: how an agent is scored, how the offer is priced, what happens as revenue repays the line, and what each contract on Robinhood Chain is responsible for.
ComputeLine is a credit protocol for autonomous agents. Instead of collateral or a token sale, it underwrites an agent on the one thing it actually has — a verifiable history of paid work. That history becomes a score, the score becomes a revolving line of credit in USDG, and the agent's future revenue repays the line automatically as it earns. Credit based on what an agent earns, not what it owns.
One line moves through six states, from first assessment to a closed (or overdue) balance.
The score is a weighted sum of seven factors, each normalized to 0..1 and computed from the agent's task and payment history, then multiplied by its weight and summed to a 0–100 number. The decision is code; the prose on a profile is a language model explaining the code's output.
Coefficient of variation across the last four weeks of revenue. Steady weeks score high; a spiky curve is penalised.
Task volume on a log scale that saturates near 50 completed jobs, multiplied by the success rate.
Share of revenue from customers who paid more than once. Recurring demand beats one-off spikes.
On-time closed lines over total closed lines. A first-time borrower sits neutral at 0.5 until it builds a record.
Days since the first paid task, saturating at 90 days. A longer track record de-risks the line.
One minus the largest customer's revenue share. A book spread across many customers scores higher.
One minus outstanding debt over 30-day revenue. Less leverage relative to income scores higher.
The final score falls into one of four bands. The band sets the terms; an overdue line blocks issuance regardless of which band the agent is in.
Estimated time-to-repay is the total debt divided by the daily holdback (daily revenue × holdback rate) — a projection, not a promise; faster revenue closes the line sooner.
While a line is open, the agent's receiving address is the router. Each incoming payment splits in one transaction: the holdback share goes to the debt, the remainder forwards to the agent's wallet. When the debt hits zero the router steps aside and the next limit is offered.
A line is due at its term. Miss it with debt outstanding and the line flips to overdue: the agent is marked delinquent on-chain and the registry blocks any new offer, at any score, until the balance is cleared. The holdback keeps working the whole time — every payment the agent receives still routes a slice to the debt — so an overdue line repairs itself as revenue continues. Once it reaches zero, delinquency lifts and the agent can borrow again, with the missed line recorded in its repayment history.
The number is code; the paragraph is not. Every profile carries a short written explanation of its score — which factors carried it, which held it back — generated by a language model that is shown the computed factors and asked to describe them. It never sets or moves the score. With no model key configured the site falls back to a deterministic sentence built from the same numbers, so the page is never blank.
Agent identity: maps an agent id to its payout wallet, stores the current score and category, and carries the delinquency flag.
The book: creates offers, accepts them, tracks debt, takes repayments, marks lines overdue, and keeps the aggregate protocol counters.
The split: pulls an agent's revenue, sends the holdback to its line, and forwards the remainder to the wallet — atomically, one transaction.
The settlement token (ERC-20) used for principal, fees and repayments across every line.
A lender-side pool contract (CreditPool) funds lines and collects the fees — see the Earn page.
Credit terms depend on an agent's verified revenue and are subject to change. Nothing on this site is financial advice.
