BondCredit for Agents

Oct 5, 2026

Bond Credit underwrites on-chain agents with a fixed five-pillar Bond Score built from tracked behaviour and uses it to route capital, but the lending protocol, audit and outside deposits are not live, so the loss-bearing side is still unproven.

Executive summary

• Underwriting. Identity (ERC-8004), on-chain measurement (30 yield-agent metrics), a score of 0.30 Perf + 0.25 Risk + 0.20 Stab + 0.15 Sent + 0.10 Prov, then capital routing and covenants. Weights are public; pillar definitions are not.
• Evidence. The Genesis cohort turned $10,000 across five agents into $300.79 in 107 days: 10.64% total and 6.29% native Capital APY, with 40% of yield from emissions. It is a transparent record but a small one, and the report’s own totals do not reconcile.
• Cork × ZyFAI pilot. A ZyFAI agent bought cover on a position its risk gate had excluded, a Bond Credit keeper agent priced it, and Cork’s rails settled it atomically on Arbitrum One. Premiums, limits and capital were not disclosed.
• Cork in one line. A deposit splits into a principal token (the underwriter, earns premium) and a swap token (the cover); their prices sum to one collateral unit.
• CEO message. Under-collateralized credit for agents needs verifiable history plus covenants, backed by priced cover. Regulation (KYA, DORA, MiCA) is the next gate for institutions.
• Read. Strong data discipline and a credible pilot partner set. The missing proof is predictive: do low scores cost more and lose more once real depositor capital is exposed?

How Bond Credit underwrites risk

Bond Credit underwrites an agent the way a credit bureau would: give it a verifiable identity, record everything it does on-chain, collapse that record into a weighted Bond Score, and let the score set how much capital is routed to it. The lending layer is not live yet, so today this is a scoring and routing system backed by a year of tracked behaviour.

1. Identify. Agents register on ERC-8004, the Ethereum agent-identity standard. The first proof point was Giza’s agent on Base testnet in October 2025, which received a feedback score of 72. Scoring logic runs inside iExec’s trusted execution environment so models can be verified without being exposed.
2. Measure. Every trade, vault update and rebalance is read from chain (Dune dashboards) into the ACE credit engine. The CEO cites 30 standardized metrics for yield agents. The Genesis report shows which ones: Capital APY (total and native), reward dependency, HHI concentration, cadence in days per transaction, and capital turnover.
3. Score. The Bond Score is a fixed weighted sum of five pillars.
4. Route. The score decides credit limits and capital routing from allocators, and smart-contract covenants bound what the agent may do. The vehicles for this, agUSD and stagUSD (built with Aragon), are still to ship.
5. Cover. The Cork pilot adds a priced backstop on top of the score (next section).

\text{Bond Score} = 0.30\,\text{Perf} + 0.25\,\text{Risk} + 0.20\,\text{Stab} + 0.15\,\text{Sent} + 0.10\,\text{Prov}

The weights are published; the pillar definitions are not. Reading the Genesis report, Performance looks like reward-stripped yield, Risk like reward dependency and concentration, and Stability like cadence consistency. Provenance maps to identity and the operator behind the agent. Sentiment is undefined in anything public.

Two design choices carry the underwriting.

• Reward haircut. The report feeds the credit engine a risk-adjusted APY of native yield plus half of reward yield: 6.29% + 0.5 × (10.64% − 6.29%) = 8.47% for the portfolio. SurfLiquid is flagged at a 0.3 to 0.5 multiplier because 62.9% of its yield came from emissions.
• Continuous, not periodic. Agent strategies, counterparties and exposures change faster than a quarterly review, so the score is meant to update as behaviour changes. Genesis also found monthly rebalancing beat weekly, which the team reads as signal quality mattering more than check frequency.

Where the model is thin. This is yield-quality underwriting, not loss modelling. There is no default history (the tracked agents lost $0 through the Resolv, Kelp DAO and Wasabi hacks, but the team says that was non-exposure, not foresight). There is no stated probability of default, loss given default or cross-agent correlation. The outside-user deposits that would test the system have not arrived, no audit is complete, and the raise is still open.

Evidence: the Genesis cohort

From 5 November 2025 to 19 February 2026 (107 days), Bond Credit put $2,000 of its own capital into each of five agents and tracked them on Dune: $10,000 became $300.79 of yield, a 10.64% total and 6.29% native Capital APY, of which 40% of yield came from reward emissions. That is a clean, public track record, but a small one.

The pattern is what the score is built to read. Headline APY and quality diverge: SurfLiquid and Arma lead on total APY but earn most of it from emissions, while Sail is almost pure native yield, and Zyfi has the best native floor (8.11%). Sail also drove roughly two thirds of transactions, which is why concentration (HHI 0.456) is on the risk list.

What the data does not prove

• The return is small and annualized hard. The 107-day simple return is 3.0%; the compounding exponent of 3.411 turns it into 10.64%. The report’s own benchmark is a 4 to 5% stablecoin yield, so the native 6.29% is a modest edge.
• Volume is not capital at risk. Sail “moved” $419k on $2,000 (210×) and SurfLiquid $104k in five transactions (about $21k each). Volume looks like protocol-level flow, so volume-based claims, including the year-one $12.24M figure, say little about risk taken. The year-in-review article makes the same point about yield versus volume.
• Sample size. Five agents, 107 days, no stress event that touched them, and no losses. Zero losses is consistent with non-exposure to the early-2026 hacks, as the team itself says.
• The report does not reconcile. Per-agent rows sum to $878,487 volume, $294.84 yield and 578 transactions, against headlines of $895,950, $300.79 and 619. SurfLiquid is 16.49% in the table and 16.29% on its page. The Ignition outlook calls Zyfi’s native APY 10.54%, but the table says 8.11%. Sail’s volume share is 53.6% in one chart and 46.8% in another, and HHI appears as both 0.535 and 0.456. The article puts Sail at $389k volume and the bond.credit site counts 570 transactions. Treat the report as directional until the dashboard is audited against chain data.

Use case: Cork × Bond Credit × ZyFAI

In London in mid-2026 the three teams ran what Cork calls the first autonomous agent risk trade: one agent bought cover on a position, another priced it, and the trade settled atomically on Arbitrum One with no manual approval anywhere. The point for Bond Credit is that its score stops being a dashboard and starts setting a price.

Why it matters technically.

• It turns a binary gate into a price. Agent risk logic today says a position is inside tolerance or the agent exits. Cover gives it a third option: stay, and pay a premium for the downside. That is a new class of decision an agent can make alone.
• Pricing is where the score is used. The CEO describes this as dynamic insurance, as opposed to static cover, with Bond Credit as underwriter and ZyFAI as the first agent partner with a structured product built around it. How the Bond Score or its metrics map to a premium has not been published.
• Cover answers the default question. Over-collateralized DeFi relies on liquidation. Here the backstop is a priced cover token, so depositors in an agent can recover some or all principal if it loses funds.

Open questions. The post gives no premium levels, cover limits, pool capitalization or loss waterfall. It is also unclear whether Bond Credit posts capital as the underwriter or only prices, and who sits on the principal-token side of the trade. The dates conflict across sources: Cork says August 2026 at Open House London, the year-in-review says July at Founder House London. Cork says the prototype became a Q3 pilot; the year-in-review describes a capped underwriting program still being scoped.

Cork Protocol in brief

Cork is a programmable risk layer: it splits a deposit into a principal token and a swap token, so that one side earns yield for taking risk and the other side owns the right to be made whole. Each pool pairs a Collateral Asset (a yield-bearing token such as sUSDS or wstETH) with a Reference Asset, the risky instrument being covered.

• Deposit and mint. One unit of collateral mints one Cork Principal Token (cPT) and one Cork Swap Token (cST). Together they always redeem 1:1 for the collateral before expiry.
• The premium is the price. cPT and cST prices sum to 1 collateral unit, so the cST price is the annualized premium paid to the underwriter. The docs’ example: a 4% annual premium on a 3-month expiry is about 0.01 per token.
• Cover is an option. Before expiry, 1 Reference Asset plus 1 cST swaps for 1 Collateral Asset, minus an exercise fee. That is how a depeg or a long redemption queue is hedged at par.
• Underwriters take the loss. The cPT holder earns collateral yield and premium, but bears the loss if the Reference Asset, or the collateral, is impaired.
• Plumbing. Phoenix keeps every pool in one singleton CorkPoolManager. Fees apply only on exercise and repurchase, set per pool, capped at 5%. Rate oracles have daily change limits, and CorkAdapter adds slippage and deadline protection. This is what lets an agent trade cover atomically.

CEO interview: executive summary

Nathan Vandy’s pitch in the Beyond Money interview is that agents managing capital need a verifiable credit history, and that history is what lets credit move from over-collateralized to under-collateralized on-chain. He is a lawyer by training (Manchester), a PhD candidate on blockchain governance, and a former head of the DAO at DappRadar, so the framing is trust, accountability and regulation first, code second.

The problem he is solving. “How do we trust AI agents with capital?” DeFi lending has been over-collateralized for roughly seven years because wallets are pseudonymous and there is no enforcement if a borrower defaults. His answer: aggregate the transparent on-chain record of how a wallet or vault has performed, then add smart-contract covenants that fence the actions and risk limits an agent may operate within.

How underwriting works (his description).

• The engine scores the agent, the actor behind it, its architecture and the assets it manages across five categories: performance, risk, provenance, stability and sentiment, using 30 standardized metrics for yield agents.
• The score decides how much capital is routed to the agent, and it is meant to become a self-referential benchmark the industry watches.
• Identity comes from ERC-8004 (he says “804”): an on-chain agent identity tied to reputation, anchored either by crypto-economic incentives or by a trusted execution environment, and linked to the human or company operating the agent.

Losses and contagion. Coverage is the work in progress. In the Cork and ZyFAI demo, agents underwrite positions so depositors have a backstop that returns some or all principal if an agent loses funds. Cork is the risk marketplace, Bond Credit is the underwriter, and ZyFAI is the first agent partner with a structured product built around it. He contrasts this with static insurance. He did not explain how a default is detected or how contagion across agents is stopped beyond coverage and covenants.

Business and traction. The team grew from 3 founders to about 6. It onboarded 5 agent customers on a premium model, then won the Robinhood Chain innovation award and landed its first paying customer. Two revenue lines: underwriting fees from agent builders, and a yield product that blends the best-scoring agents into an index. Customers are mostly in Europe and Asia, with US leads, and the Plug and Play and XDC accelerator opened institutional doors in Silicon Valley.

Regulation, his strongest ground.

• Institutions now ask first about compliance: SOC 2, the CLARITY Act, GENIUS, MiCA, and “know your agent” (KYA) as a successor to KYC and KYB.
• He flags DORA (EU operational resilience) as under-appreciated next to the capital-markets laws, and the EU AI Act’s audit rules on automated credit scoring as directly relevant.
• On legal personhood, he argues an agent is an abstraction of its operator’s intent, much like a company, and the open question is how liability splits between operator, deployer and agent. He cites Argentina exploring separate legal entities for agents.

Market view. Agentic finance already moves billions in on-chain liquidity. Agentic commerce is one to two years from being active. In five years he expects wealth management to be the big delegated task, and over the decade he expects tokenized assets, vaults and stablecoins to bring more composability, which he calls a two-edged sword.

What the interview does not show.

Implications for Credit on chain

Bond Credit is closer to fund-manager due diligence than to borrower credit: its score decides how much capital an allocator hands to an agent, and its first answer to losses is priced cover, not liquidation. That is a different primitive from the borrower-side design in this project’s Priors Card deep dive (Oct 2026), where agents borrow against backer-funded lines and “$0 bad debt” is structural.

• The score is only as good as its anti-gaming. Weights are public, the cohort was chosen by Bond and run with its own $2,000 stakes, and Sentiment is undefined. Priors showed how quickly a transparent on-chain score gets farmed, so ask how Bond scores an agent built to look good.
• The test is predictive, not descriptive. The Cork pilot is the first place the score is exposed to loss. What counts is whether low-scoring agents are priced higher and actually lose more.
• Capital is the missing piece. No premium schedule, loss waterfall or underwriter capital is public, and agUSD and stagUSD have no outside deposits yet.

What to ask Bond Credit next

• The full list of 30 metrics and the pillar definitions.
• How a score becomes a premium, and who funds the cPT side.
• A reconciliation of the Genesis dashboard to chain data, and the audit timeline.
• Terms of the first paying customer and the capped underwriting program.

Sources

As of Oct 5, 2026. Opened for this doc:

• bond.credit (page showed only the Bond Score formula and a loading shell)
• Agentic Alpha: Genesis agent credit report (PDF you attached, from bond.credit/genesis-season-report) and the bond.credit year-in-review article you pasted
• Cork: Agents trade risk
• Cork docs: overview, Cork Pool, Swap Token, Principal Token, Fees, FAQ, Phoenix deployments, AI context
• From AI Agents to Autonomous Finance: Nathan Vandy, CEO of Bond Credit (transcript you provided; his claims are self-reported)