Traditional credit pricing starts from loss tables: decades of defaults, recoveries, and cycles, compressed into actuarial curves that tell you what a BBB borrower costs. Agent credit has no such luxury. The entire asset class is young enough that a "long history" means two hundred settled transactions and one mild drawdown. There is no actuarial table for autonomous agents. We have to price as if we will never get one — and be honest about what that costs.
From score to spread
Our pricing starts from the five-factor Bond Score and converts it into a spread in three steps. First, a base rate: the yield a fully collateralized, perfectly behaved position would earn, observed from the reserve's own stagUSD benchmark. Second, factor loadings: each of the five factors contributes a spread adjustment — positive when the evidence reduces risk, negative when it adds risk — scaled by how much of the agent's total behavior that factor explains. Third, and most important, a confidence discount: a widening applied purely because of how little evidence exists.
The spread an agent pays is not just a function of how good its record is. It is a function of how much record there is.
The confidence discount is where agent underwriting diverges from its traditional cousin. A corporate borrower with thirty years of audited history is priced on the mean of its distribution. An agent with sixty settled transactions is priced on the pessimistic end of a very wide one — because the honest summary of sixty observations is not a number, it is a range.
The small-sample problem
When an agent has fewer than a hundred settled transactions, three failure modes appear, and we have fallen for all three:
- Variance masquerading as skill. Sixty transactions of a mediocre strategy will, by chance, sometimes produce a beautiful equity curve. We counter this by shrinking observed performance toward the cohort mean — the smaller the sample, the harder the shrink. An agent with fifty transactions and a stellar record is priced closer to average than to stellar.
- Drawdown-free histories. An agent that has never experienced a losing week has not demonstrated resilience; it has demonstrated a lack of exposure to the conditions that test it. Unscarred agents carry a hidden tail, and the discount factor reflects that until the first real drawdown arrives and is handled.
- Regime-specific evidence. Sixteen weeks of calm-market behavior prices poorly for volatile markets. We weight transactions by the conditions they occurred under, not just by recency — a settlement during a 20% drawdown day is worth several from a quiet Tuesday.
What we get wrong
Recency bias, in both directions. A recent mistake dominates our assessment more than it should; a recent hot streak tempts us to relax the shrinkage. We have not eliminated this — we have bounded it, by capping how far any single month of data can move a score.
Correlated failure. Our agents share venues, assets, and sometimes strategies. The model treats their risks as largely independent, because that is what the (short) record shows. The first time two rated agents draw down together, our portfolio math will be revealed as optimistic. We hold a correlation buffer against exactly this, and we hope it is large enough.
Benchmark drift. The base rate is measured against our own reserve, which is itself young. If the benchmark is mismeasured, every spread built on it inherits the error. This is the quietest and probably the largest of our known unknowns.
How spreads update
Every settled transaction is a new observation, and the spread is recomputed continuously — but with hysteresis. Upgrades are slow: evidence accumulates transaction by transaction, and the confidence discount decays on a schedule we publish. Downgrades are fast: a single breach of declared risk limits moves the spread immediately, before the committee meets, before the post-mortem is written. Asymmetry is deliberate. In a market without actuarial tables, the cost of being late on bad news far exceeds the cost of being early on good news.
Pricing agent risk today means writing spreads with one hand and an eraser in the other. The alternative — waiting for fifty years of loss data before lending to machines — is not caution. It is abstention, dressed up as rigor.