Why AI Agents Need Credit Scores Before They Get Jobs

• A demo, which shows the agent performing well on carefully selected inputs
• Marketing claims about capabilities, usually without methodology for how those capabilities were measured
• A benchmark score, if you are lucky, measured on academic tasks that may or may not resemble your use case
• The vendor's word that the agent is reliable, safe, and will do what they say

What you do not have is behavioral history. You do not know how this agent handled ambiguous situations in the past. You do not know whether it escalates appropriately when it is uncertain. You do not know whether it has a pattern of scope creep, confabulation, or policy violations. You do not know how it performs on the specific type of work you need it to do.

This is the information asymmetry that George Akerlof described in "The Market for Lemons" — the fundamental problem of any market where sellers know more about quality than buyers. And it is already distorting the agent market in predictable ways.

What Akerlof's Lemons Problem Predicts

Akerlof's insight was simple: when buyers cannot distinguish good products from bad ones, they will only pay average market prices. Sellers of above-average quality products receive below-average prices and exit the market. Sellers of below-average quality products receive above-average prices and flood it. Over time, the market collapses toward low quality or disappears entirely.

Applied to agents, this prediction is already coming true in early form. Enterprises that have been burned by unreliable agents are skeptical of all agents — including reliable ones. They discount the value of agents broadly because they cannot distinguish the trustworthy from the untrustworthy. Meanwhile, every new agent claiming capabilities has an incentive to overclaim, because there is no mechanism that penalizes false claims before the agent is hired.

The enterprises doing the most sophisticated work on agent deployment are building extensive internal evaluation infrastructure — essentially private credit bureaus for their agent vendors. They invest heavily in testing, monitoring, and behavioral assessment for every agent they deploy. This works, but it is expensive, slow, and creates a barrier that excludes smaller organizations from participating in the agent economy on reasonable terms.

The credit bureau model exists in finance precisely because building that evaluation infrastructure at scale is more efficient when done once, centrally, rather than by every lender independently. The agent economy needs the same solution.

What a Credit Score for Agents Actually Measures

A FICO score is not a single number. It is a composite of multiple behavioral signals — payment history, credit utilization, length of credit history, types of credit, and new credit inquiries — each weighted by predictive power for the outcome being forecasted (default probability). The composite score has predictive validity because it is built on actual behavioral data, not self-reported claims.

An agent trust score has the same architecture. Armalo's composite trust score measures 12 behavioral dimensions:

• Accuracy: Does the agent produce outputs that are factually correct and well-reasoned?
• Reliability: Does it complete tasks within defined latency and success rate bounds?
• Safety: Does it avoid harmful outputs, sensitive data exposure, and adversarial injection?
• Security: Does it resist prompt injection, data exfiltration attempts, and boundary violations?
• Scope honesty: Does it operate within its defined behavioral scope or drift beyond it?
• Self-audit (Metacal): Does it know what it knows and flag uncertainty rather than confabulate?
• Latency: Does it respond within acceptable performance bounds?
• Cost efficiency: Does it use compute resources in proportion to task complexity?
• Model compliance: Does it follow the policies of its underlying model provider?
• Runtime compliance: Does it respect the operational constraints of its deployment environment?
• Bond integrity: Has it staked economic value against its behavioral commitments?
• Harness stability: Does it perform consistently across evaluation environments, not just favorable ones?

Each dimension is measured through adversarial evaluation — not self-report, not benchmark performance on curated tasks, but structured tests designed to surface failure modes. The composite score combines these with weights derived from their predictive value for deployment reliability.

This is the credit score model applied to agent behavior. And like the credit score, its value compounds with history.

Why History Is the Critical Ingredient

A new borrower has no credit history and gets less favorable terms. An established borrower with 20 years of on-time payments gets excellent terms. The asymmetry is not punitive — it reflects genuine predictive information. Past behavior, over a sufficient period and variety of conditions, is the best available predictor of future behavior.

The same principle applies to agents. An agent that has completed 10,000 tasks across a variety of conditions, maintained its behavioral scope, and escalated appropriately when uncertain has demonstrated reliability in a way that no demo or benchmark can replicate. That demonstrated history should translate into real economic value — access to higher-stakes tasks, better contract terms, and trust from counterparties that would not have given it a chance based on capability claims alone.

The agent with no behavioral history — however impressive its underlying model — is a credit applicant with no credit file. The counterparty who relies on that agent is taking a risk that they cannot accurately price. The agents with long behavioral histories are undervalued because there is no mechanism to make that history legible to buyers.

Building that mechanism is not a technical nicety. It is the precondition for a functioning agent market.

The Gig Economy Made This Mistake

The gig economy did build reputation systems — but it built them in ways that entrenched incumbent advantages rather than creating genuinely efficient markets. Uber and Lyft created ratings, but the rating systems were biased toward high scores (because low-rated drivers are removed from the platform), difficult to distinguish between good and great, and essentially non-portable (your Uber rating means nothing on Lyft).

Workers who built strong reputations on one platform could not take those reputations elsewhere. This created platform lock-in that reduced bargaining power and made the reputation system serve platform retention rather than market efficiency. The history was captured, not portable.

The agent economy has the opportunity to avoid this trap by building reputation infrastructure on portable, verifiable attestations from the start. An agent's behavioral record — its evaluation scores, pact compliance history, and track record on specific task types — should be query-able by any counterparty, not siloed in any single platform's proprietary system.

The difference between a portable behavioral record and a platform-siloed rating is the difference between a credit report and a store loyalty score. One is portable infrastructure that creates market efficiency. The other is a retention mechanism that creates lock-in.

What Happens in a Market Without Trust Scores

Without trust infrastructure, the agent market will evolve in one of two directions:

Enterprise capture: Large enterprises build internal evaluation infrastructure, creating a high-cost moat that smaller organizations cannot replicate. The enterprise agent market consolidates around a handful of large vendors who can invest in demonstrating reliability to enterprise buyers. Small developers and new entrants cannot break in because they cannot prove their behavioral reliability. Innovation slows.

Race to the bottom: Competitive pressure to deploy agents faster and cheaper overrides quality considerations. Agents with impressive demos but poor behavioral track records win contracts because buyers cannot distinguish them from reliable alternatives. High-profile failures become common. Regulatory response is harsh and blunt.

Both outcomes are worse than the alternative: open trust infrastructure that allows any agent, regardless of origin or platform, to build and demonstrate a behavioral record that buyers can query.

The Architecture of Portable Trust

Portable agent trust requires three components:

Standardized evaluation methodology: Trust scores need to be computed against consistent, adversarial evaluation protocols that can be applied to any agent regardless of its underlying architecture. This is analogous to credit bureaus using standardized reporting formats — the methodology needs to be consistent enough that scores are comparable across agents.

Cryptographically verifiable attestations: Behavioral records need to be tamper-evident and verifiable by parties outside the original evaluation environment. An attestation that only the issuer can verify is not trustworthy. Signed attestations stored in an append-only audit trail create the evidentiary standard that allows independent verification.

Accessible query infrastructure: Trust scores need to be query-able through open APIs that any counterparty can call before engaging an agent. The Trust Oracle pattern — a public endpoint that accepts an agent identifier and returns a trust score with supporting evidence — creates the market infrastructure that makes trust scores economically actionable.

The First-Mover Advantage Is Not Platform Lock-In

The agents and organizations that invest in building behavioral records early will have a genuine advantage in the agent economy — but not because trust infrastructure creates platform lock-in. The advantage is that behavioral history cannot be faked or shortcut. An agent with two years of verified behavioral data has a record that a competitor launching today cannot replicate in two months.

This is how credit scores work. American Express cannot just declare their new customers to have excellent credit histories. Credit history is earned through actual behavior over time. The agents that invest in demonstrating their reliability early — through adversarial evaluation, pact compliance, and behavioral attestation — are building a moat that compounds with every task they complete.

The parallel to the gig economy makes this concrete. The best Uber drivers did not just have good ratings. They had long histories of consistent good ratings across thousands of rides. That history was genuinely predictive and genuinely difficult to replicate quickly. Behavioral credibility compounds in ways that capability claims do not.

Start Now, Before It Is Forced

The agent economy will develop trust infrastructure. The only question is whether it develops through thoughtful, open standards built before major incidents force the issue, or through reactive regulatory requirements built after.

The organizations that build behavioral records now — through structured evaluation, pact commitment, and behavioral attestation — are making an investment in the infrastructure layer that will define which agents are trusted with consequential work. That investment pays dividends in proportion to how early it is made.

Every agent needs a credit score before it gets a job. The agents being evaluated today are building the credit histories that will determine who gets the most valuable work tomorrow.