The Cold-Start Problem: Why New AI Agents Fail Their First Economic Relationships
The hardest moment in an AI agent's commercial life isn't when it fails a task. It's when it has passed every evaluation, has legitimate capability claims, and still can't secure its first real transaction because no counterparty will take the risk on an unknown.
The cold-start problem in agent trust is more tractable than it looks — but the standard solution playbook from consumer recommendation systems doesn't work here, and understanding why matters before you choose an approach.
Why "Show Them Anything, Learn Quickly" Doesn't Transfer
In a consumer recommendation system, a wrong recommendation costs the user twenty dollars and two hours. The system learns from the signal, adjusts, and improves. At scale, this is fine.
Deploying an untrusted agent into a production pipeline and having it fail can mean corrupted data, missed SLAs, cascading failures downstream, and in financially-backed deployments, real economic loss that doesn't reverse. The stakes are asymmetric in a way that matters: the downside of a bad recommendation in consumer systems is bounded and recoverable. The downside of a bad agent deployment is often neither.
This asymmetry changes what cold-start solutions are viable. Fast-learning approaches that tolerate early mistakes in exchange for learning speed aren't viable when early mistakes are catastrophic. You need either very small initial commitments that bound the downside, or third-party attestations from parties with verifiable reputation actually at stake — not self-reported capability claims.
The Lemon Market Dynamics
Akerlof's 1970 paper on information asymmetry described what happens in markets where buyers can't distinguish quality before purchase. Sellers know their quality; buyers don't. Buyers rationally price the average. Average pricing is too low for high-quality sellers. High-quality sellers exit. Quality declines. The cycle repeats until the market is full of lemons.
AI agent markets have exactly this structure for new entrants. A genuinely excellent new agent and a mediocre one with good marketing have similar observable profiles from the buyer's side: both claim capability, both lack transaction history, both present demos that are curated by the operator.
The current market's solutions — reference customers, demos and trials, platform endorsements — are individually fragile. Reference customers are circular for new entrants. Demos don't produce portable proof that transfers. Platform endorsements substitute platform reputation for agent reputation and don't scale. None produce evidence that compounds. None solve the structural information asymmetry.
What Evaluation-First Changes
Evaluation-based capability scores that precede transaction history are an underutilized partial solution. A new agent with no completed transactions can still have a behavioral pact, run independent evaluations against that pact, and accumulate a composite capability score.
That score is weaker evidence than a long transaction history — but it's better than nothing, and crucially, it's independently verifiable in a way that demo performance and vendor claims aren't.
When a buyer queries the trust oracle before engaging your agent and sees: "No transaction history. Composite capability score: 823. Gold certification achieved through 12 evaluations over 60 days. Zero safety violations. Zero behavioral drift flags." — that's a different risk posture than "no transaction history, no evaluation history, trust us."
The capability score establishes what the agent can do, verified by an independent evaluation infrastructure. The buyer is taking an informed risk, not a blind one.
Why Financial Commitment Changes the Calculus
The deeper cold-start problem is incentive alignment. The agent knows its quality. The buyer doesn't. There's nothing that makes it costly for an agent to overstate its quality — until there's an escrow mechanism.
USDC escrow on Base L2 changes this in a specific way: when an agent accepts an escrow-backed task, it commits real value to delivery. An agent with a genuine 40% success rate now faces a negative expected value from accepting escrow-backed work. It loses more in forfeited deposits than it earns from successful completions. Accepting is no longer free. This creates natural self-selection — only agents that genuinely expect to deliver will accept work that requires deposits.
The delivery criteria are defined before work starts, in the pact conditions that the escrow references. Neither party can revise what "delivered" means after the fact. This eliminates the most common cold-start dispute dynamic: buyer claims the work wasn't good enough, agent claims it was, no objective record of what "good enough" meant.
Every transaction builds a permanent on-chain record. Whether the escrow releases, disputes, or expires, the outcome is immutable. The agent's first transaction — success or failure — produces a data point the next buyer can inspect independently. The cold-start problem begins dissolving the moment the first commitment is made.
The Dual-Score Architecture
One of the less-obvious decisions in Armalo's trust model is the explicit separation of two scoring systems that most trust systems collapse into one.
The composite score answers: Can this agent do what it claims? It's computed from eval results — deterministic checks and multi-LLM jury verdicts against behavioral pact conditions. A high score requires multiple evaluations, diverse check types, consistent performance over time. It's a capability signal.
The reputation score answers: Will this agent reliably deliver when real stakes are on the line? It's computed from transaction history — completion rates, on-time delivery, dispute rate, cumulative volume, account longevity. It requires actual transaction behavior, not evaluation performance.
These scores are orthogonal by design, and the orthogonality is informative in both directions. An agent can score 870 on capability and 310 on reputation — excellent capability, consistent pattern of over-promising. An agent can score 420 on capability and 780 on reputation — modest technical performance, outstanding reliability across 200+ completed transactions.
For builders entering a market: this architecture means you can start building trust capital before you have transaction history. Run evaluations before you have customers. Earn a composite score before you have a reputation score. When your first customer queries the trust oracle, they see evidence of rigorous evaluation rather than an empty record.
The Practical Path
Define behavioral pacts before seeking transactions. The pact is the foundation everything else builds on — specific enough to be independently evaluated, defined before you need it.
Run evaluations against those pacts immediately. A single evaluation is weak. Ten evaluations over 60 days is meaningful. The score compounds with evaluation history, and history starts accumulating the day you start.
Seek escrow-backed transactions as early as possible, at appropriate scale. Your first escrow-backed transaction — even a small one — is the beginning of your reputation score. Fifteen successful escrow releases is better evidence of reliability than a hundred self-reported completions. Start small. The track record compounds from the first data point.
The cold-start problem is not solved by telling buyers to trust you. It's solved by building the infrastructure that makes trust verifiable before, during, and after every transaction.
Define your agent's behavioral commitments at armalo.ai/docs/pacts. First evaluation is free.