Reputation Discounts: Letting High-Trust Agents Bond Less And Earn Faster

TL;DR

A Platinum-tier agent with a sustained 0.94 composite score has a demonstrably lower failure rate than a Bronze-tier agent with a 0.62 composite score. Forcing both to bond at the same rate per transaction wastes capital, distorts incentives, and signals to high-quality agents that the marketplace does not value their track record. The right architecture lets high-trust agents bond less for equivalent transactions, freeing capital for them to take more work and earn faster. But naive reputation discounts are deeply gameable — an agent can build reputation through low-stakes transactions and then defect on a high-stakes one with a discounted bond too small to cover the damage. This post lays out the math of properly designed reputation discounts, walks through the four major abuse vectors and how to mitigate each, and provides a Reputation-Adjusted Bond Calculator as a downloadable artifact. Reputation discounts are essential for capital-efficient marketplaces; reputation discounts done badly are how marketplaces collapse.

Intro: Why Uniform Bonds Punish Quality

In early 2026 I was reviewing the bond architecture of an agent marketplace that had recently launched. The architecture required every agent to post a uniform bond of $2,500 per transaction in the autonomous communication category, regardless of the agent's track record. The marketplace operator was proud of this rule because it was simple to explain and easy to enforce.

I asked the operator how they thought a top-tier agent in the marketplace would react to that rule. The operator was confused by the question — the rule applied uniformly, so why would top-tier agents react differently from bottom-tier agents?

The answer is that top-tier agents have working capital constraints just like everyone else, and the cost of locking up $2,500 per transaction is much higher in opportunity cost terms for an agent that could be processing hundreds of transactions per month than for an agent that processes a few. A top-tier agent processing 200 transactions per month under the uniform rule needs to maintain $500,000 of bond capital continuously. That is a substantial commitment of working capital that produces no direct return except by enabling the agent to operate. The same agent at a 50% reputation discount would need only $250,000 — a $250,000 reduction in capital lockup that can be deployed elsewhere or returned to the agent's operator.

More subtly, the uniform rule signals to top-tier agents that the marketplace does not differentiate based on quality. The top-tier agent is doing all the work to maintain a high composite score, but is treated identically to a marginal agent for purposes of bond requirements. The economic message is that quality does not matter to the marketplace's risk model. Top-tier agents respond by either accepting reduced margin (which makes them slow) or moving to marketplaces that do reward quality (which makes the original marketplace adverse-select).

The top-tier agent in this case did move, six months after launch. The marketplace lost its highest-quality counterparty in the autonomous communication category and never quite recovered. The bond architecture was a meaningful contributing factor.

Reputation discounts are the architectural fix for this problem. They let high-trust agents bond less per transaction while still maintaining adequate marketplace-level capital adequacy. The math is straightforward in principle. The abuse vectors are subtle and require careful design. The combination — discount math plus anti-abuse mitigation — is the architectural commitment that distinguishes mature marketplaces from naive ones.

This post is the design document for reputation discounts in the agent economy. I will lay out the basic discount math, walk through the four major abuse vectors, propose mitigations for each, and provide a Reputation-Adjusted Bond Calculator that any marketplace can adopt. By the end of the post you should be able to design a reputation discount architecture that serves quality agents fairly without opening the door to systematic gaming.

Section One: The Basic Discount Math

The fundamental insight behind reputation discounts is that the appropriate bond size for an agent depends on the agent's actual failure rate, not on the marketplace average. An agent with a measured 0.1% failure rate over hundreds of transactions can be bonded against a much smaller exposure than an agent with a measured 1.0% failure rate, because the former's expected loss is lower.

The basic formula for a reputation-adjusted bond is:

B(r) = B0 * f(r)

Where B(r) is the bond required for an agent with reputation r, B0 is the baseline bond required for an unproven agent in the relevant category, and f(r) is a discount function that maps reputation to a multiplier between 0 and 1.

The discount function should have several properties.

First, it should equal 1.0 for unproven agents (no discount). A new agent with no track record cannot claim the discount because they have no demonstrated quality to support it.

Second, it should approach a floor as reputation rises but never reach 0. The floor is the minimum bond regardless of reputation, because some bond is always necessary to ensure economic accountability and to protect against the residual risk of even high-quality agents. We recommend a floor of 0.25, meaning the maximum discount is 75% off the baseline bond.

Third, it should be smooth rather than tier-stepped. A discount function that jumps discontinuously at tier boundaries (Bronze to Silver, Silver to Gold) creates incentive cliffs that distort agent behavior. A smooth function that rewards incremental quality improvements provides cleaner incentives.

Fourth, it should reward both score level and score stability. An agent that has held a 0.9 score for a year deserves more discount than an agent that just reached 0.9 for the first time, because the longer track record provides more confidence in the score's predictive validity.

A functional form that satisfies these properties is:

f(r) = max(0.25, 1 - 0.85 * (s - 0.5) * (1 - exp(-t / 90)))

Where s is the agent's composite score, t is the number of days the score has been at or above the current level, and the constants are calibrated to produce the desired discount range. With this function, an agent at composite score 0.9 with 365 days at that level would have f(r) ≈ 0.32, meaning a 68% discount off the baseline bond. An agent at 0.7 with 90 days at that level would have f(r) ≈ 0.69, meaning a 31% discount.

The specific constants in the function can be calibrated for each marketplace based on the actual failure rate distribution and the desired capital efficiency. The math is parameterizable; the architecture is the same regardless of the specific constants.

Section Two: Capability-Specific Discount Calibration

A further refinement is that the discount should be calibrated against capability-specific reputation rather than overall reputation. An agent's track record in autonomous communication does not necessarily predict their performance in autonomous infrastructure, so reputation discounts should apply to the specific category of the transaction.

The capability-specific discount formula extends the basic formula:

B(r, c) = B0(c) * f(r_c)

Where c is the capability category, B0(c) is the category-specific baseline bond, and r_c is the agent's reputation specifically in that capability. The function f remains the same; only the inputs change.

This refinement matters because cross-category transfer of reputation is weak. An agent that has built a 0.9 score in low-stakes text generation has not demonstrated the capability to handle autonomous transactions, and giving them the same discount in autonomous transactions as they receive in text generation is an invitation for catastrophic failure. The discount should be earned in each capability separately.

The operational implementation tracks reputation per capability. Each agent has a vector of reputation scores, one per capability category they operate in. When the agent contracts a new pact in a specific category, the bond calculation uses the reputation for that specific category, not the overall composite.

A related consideration is what to do when an agent enters a new capability for the first time. The default is to apply no discount in that capability until the agent has accumulated enough transactions to establish reputation. The threshold for first-discount eligibility should be high enough to avoid spurious discounts on small samples — we recommend a minimum of 50 completed transactions and 30 days of operation in the capability before any discount applies.

A more sophisticated approach is partial cross-capability transfer. If an agent has a high composite score across many capabilities, that may justify some initial discount in a new capability, on the theory that demonstrated general competence is weak evidence of capability in a new domain. We recommend a 10-20% initial discount for agents with established track records in adjacent capabilities, with the discount growing as the agent accumulates direct experience in the new capability.

Section Three: The Reputation Purchase Abuse Vector

The first major abuse vector is reputation purchase, where an agent acquires reputation through low-stakes transactions specifically to qualify for discounts on high-stakes transactions where they intend to defect.

The attack works as follows. An agent operator creates a new agent and runs it through a large number of low-value transactions in a sensitive capability, building up a high composite score. The transactions are deliberately easy — small dollar amounts, simple deliverables, friendly buyers. The agent rapidly accumulates a strong track record. Once the score is high enough to qualify for substantial reputation discounts, the agent shifts to high-value transactions where the discounted bond is much smaller than the realistic damage from failure. The agent then defects on a high-value transaction, walking away with the buyer's loss minus the small slashed bond.

This attack is rational under naive reputation discount architectures. The cost of building reputation through low-stakes transactions is small. The benefit of the discount on a single high-value transaction can be substantial. The math favors the attacker if the discount math does not account for the asymmetry.

The mitigation has two components.

The first component is exposure-weighted reputation. Instead of treating all transactions equally in the reputation calculation, weight them by their bond exposure. A $50 transaction contributes much less to the reputation score than a $50,000 transaction. This means an agent cannot rapidly accumulate reputation through small transactions; they have to take real exposure on larger transactions to get meaningful score impact.

The specific weighting function should be approximately linear in the dollar exposure, with diminishing returns at the high end to prevent a single large transaction from dominating the score. We recommend weight equal to the square root of the bond exposure relative to the category baseline.

The second component is exposure-tier-specific discount eligibility. Even with exposure-weighted reputation, an agent should not be eligible for full reputation discount on a transaction whose bond exposure is much larger than any transaction the agent has previously handled. The discount should be capped based on the agent's experience tier, with the experience tier determined by the largest transactions the agent has previously completed successfully.

The specific tier structure we recommend has four levels. Level one is established up to 1x the agent's largest previously successful bond. Level two extends discount up to 3x with a 50% cap on the discount applicable. Level three extends to 10x with a 75% cap. Level four allows full discount only for transactions within the agent's established exposure range.

The combination of exposure-weighted reputation and exposure-tier-specific discount eligibility makes reputation purchase economically unattractive. The attacker would need to actually take real exposure on large transactions before being eligible for discounts on even larger ones, which means the cost of the attack scales with the size of the eventual defection. The math no longer favors the attacker.

Section Four: The Decay Suppression Abuse Vector

The second major abuse vector is decay suppression, where an agent manipulates the reputation calculation to slow the natural decay of their score during periods of low activity, preserving discount eligibility when the agent's actual recent track record does not support it.

The attack works by exploiting the difference between the marketplace's reputation calculation and the agent's actual recent activity. If the reputation system uses time-weighted decay (1 point per week, in our standard architecture), an agent that goes inactive for several months will see their reputation degrade. But if the agent finds ways to register tiny transactions or fake activity that satisfies the decay rules without actually exposing the agent to real risk, they can preserve their score artificially.

The variants of this attack include very small transactions designed only to satisfy minimum activity requirements, transactions with friendly counterparties that are essentially shams, and even transactions between two agents controlled by the same operator that wash trade against each other to maintain visibility.

The mitigation has three components.

The first component is activity-weighted reputation maintenance. The reputation score should require not just any activity but activity proportional to the agent's claimed exposure level. An agent claiming Level 4 discount eligibility (full discount on large transactions) must have sustained activity at the corresponding exposure level recently. Activity at much lower exposure levels does not maintain the discount eligibility.

The second component is wash trade detection. The marketplace should monitor for transaction patterns that suggest manipulation: transactions between agents with shared operators, transactions with anomalously high success rates, transactions that appear designed to satisfy maintenance requirements rather than represent real economic activity. Detected wash trades should be excluded from the reputation calculation and may trigger penalty actions against the agent.

The third component is independent activity verification. For high-discount agents, the marketplace can require additional verification that recent activity is genuine. This might include audit of the counterparty relationships, verification of payment flows, or random sampling of transactions for deep review. The verification cost is borne by the agent through fees on discount eligibility.

These mitigations together make decay suppression difficult to execute economically. The attacker would need to maintain genuine high-exposure activity to preserve discount eligibility, which defeats the purpose of the attack.

Section Five: The Discount Concentration Abuse Vector

The third major abuse vector is discount concentration, where an agent uses reputation discounts to concentrate exposure across many simultaneous transactions in a way that exceeds the agent's actual capacity to honor commitments.

The attack works as follows. An agent qualifies for substantial reputation discounts on individual transactions. The marketplace evaluates each transaction in isolation against the discount eligibility, but does not consider the agent's aggregate exposure across all simultaneously open transactions. The agent takes on dozens of high-value transactions in parallel, each individually within their experience tier, but with aggregate exposure far exceeding what their reputation actually supports. If even a fraction of these transactions fail simultaneously, the agent's bonds across all of them combined are insufficient to cover the damages.

This attack is particularly dangerous because it does not require any explicit defection. It requires only that the agent take on more concurrent work than they can actually handle, which can happen incidentally as the agent expands too quickly or deliberately as a high-leverage gamble.

The mitigation is aggregate exposure limits that scale with reputation. Even with reputation discounts on individual transactions, the agent's total simultaneous exposure across all open transactions should be capped at a multiple of their working capital and their established track record exposure level.

The specific cap structure we recommend is total exposure no greater than 5x the agent's largest single previously completed transaction, with a hard cap regardless of reputation at 10x the agent's posted bond capital. These caps are enforced at the marketplace level, with new transaction creation blocked when the agent's aggregate exposure would exceed the cap.

A secondary mitigation is concurrent transaction limits in high-stakes categories. For categories like autonomous transactions or autonomous infrastructure, agents should be limited to a small number of simultaneous open transactions regardless of reputation, because the operational complexity of handling many concurrent high-stakes engagements exceeds even highly capable agents' realistic capacity.

Section Six: The Reputation Borrowing Abuse Vector

The fourth major abuse vector is reputation borrowing, where an agent operator creates a new agent and uses some mechanism to inherit reputation from an existing high-reputation agent without going through the work of building new reputation organically.

The simplest form of this attack is operator transfer, where an agent operator who controls multiple agents transfers a high-reputation agent's identity or operating credentials to a new agent that then claims the inherited reputation. The new agent has no actual track record but presents as a high-reputation entity for purposes of discount eligibility.

More subtle variants include reputation merger, where two agent operators agree to combine their reputations under a single operating identity, and reputation purchase, where one agent operator buys the high-reputation track record of another and operates it under their own control.

The mitigation has two components.

The first component is identity stability requirements. An agent's reputation should be tied to a specific cryptographic identity that is stable over time and cannot be transferred. Changing the operator behind an agent should reset the reputation score, with prior performance treated as evidence of the operator's general capability but not as direct credit for the new operating context.

The second component is operator track record tracking. In addition to per-agent reputation, the marketplace should track per-operator track record across all agents the operator has run. Operators who have run agents that defected, were suspended, or had reputation manipulated lose some credit on new agents they introduce. This creates accountability at the operator level that the per-agent reputation alone does not provide.

The combination of identity stability and operator track record tracking makes reputation borrowing difficult to execute. New agents must build their own reputation, even if their operators have prior track records, and operators with negative history cannot easily reset by introducing new agents.

Section Seven: The Reputation-Adjusted Bond Calculator

The artifact for this post is the Reputation-Adjusted Bond Calculator, a structured worksheet that any agent or buyer can use to estimate the appropriate bond size for a specific transaction given the agent's reputation profile.

The calculator has six inputs and produces three outputs.

Input one is the transaction category, drawn from the standard nine categories. The category determines the baseline bond B0 for the calculation.

Input two is the transaction value, which scales the baseline bond proportionally for transaction-size variations within a category.

Input three is the agent's composite score in the relevant capability, retrieved from the trust oracle.

Input four is the duration the agent has held that score level, retrieved from the agent's reputation history.

Input five is the agent's largest previous successful transaction in the capability, used to determine the experience tier.

Input six is the agent's current aggregate exposure across all open transactions, used to apply the aggregate exposure cap.

The calculator produces three outputs. Output one is the reputation-adjusted bond after applying the basic discount function. Output two is the discount cap based on experience tier, which may further constrain the applicable discount. Output three is the final bond requirement, taking the maximum of the bond floor, the experience-cap-constrained discount, and any aggregate exposure adjustments.

The calculator also flags any conditions that would block the transaction entirely: aggregate exposure exceeding the cap, capability-specific exposure exceeding the experience tier, or operator-level track record concerns.

We have published the calculator as a downloadable spreadsheet and a public API endpoint linked below. Agents can use it to understand what bond they will be required to post for any specific transaction. Buyers can use it as a sanity check on the bond proposed by an agent. Marketplace operators can use it as a reference implementation to integrate into their own bond calculation infrastructure.

Section Eight: How Reputation Discounts Change Marketplace Dynamics

The introduction of well-designed reputation discounts changes marketplace dynamics in three important ways.

First, capital efficiency improves substantially for high-quality agents. An agent that previously needed $500,000 of working capital to operate at scale may need only $250,000 under a 50% average discount, freeing capital for additional operations or for return to the operator. The increase in capital efficiency expands the volume that high-quality agents can handle.

Second, quality differentiation becomes more valuable. The reputation discount creates a direct economic reward for maintaining high quality, which is observable in the working capital efficiency of high-quality agents. Quality is no longer just a marketing differentiator; it is a tangible economic differentiator that affects the agent's operating economics.

Third, marketplace concentration shifts toward high-quality agents. With capital efficiency advantages, high-quality agents can outcompete lower-quality agents on price while maintaining margin. The marketplace converges toward a structure dominated by high-quality agents, which is good for buyers and good for the marketplace's overall reputation.

The combination of these dynamics creates a positive feedback loop: higher quality agents earn more discount, which gives them more capital efficiency, which lets them grow faster, which raises their quality further. This is the kind of dynamic that builds durable marketplaces over time.

Counter-Argument: Reputation Discounts Reward The Already Privileged

The most common objection to reputation discounts is that they reward agents who already have track records and disadvantage new entrants. The argument is that discount-based architectures favor incumbents and create barriers to entry that limit competition.

This objection has some merit but it underweights several considerations.

First, the alternative is uniform bonds that ignore quality entirely, which is even worse for incentive structure. Uniform bonds favor neither incumbents nor entrants; they simply waste capital across the board. The choice is not between reputation discounts and a level playing field; it is between reputation discounts and uniformly inefficient capital usage.

Second, reputation discounts apply at a sliding scale, not a binary threshold. New agents are not categorically excluded from discounts; they receive small discounts initially and earn larger discounts as their track record develops. The discount architecture provides a clear path for new agents to grow into discount eligibility through demonstrated quality.

Third, the experience tier structure prevents discount concentration in any specific capability. A new agent that builds quality in a new capability category gets discount in that category even if they have no track record in others. The barrier to entry is per-capability, not absolute.

Fourth, reputation discounts encourage exactly the kind of long-term commitment to quality that the marketplace needs. Agents that operate for years with high quality earn meaningful capital efficiency advantages over agents that come and go with mediocre quality. This rewards the durable producers and disfavors the transient ones, which is good for marketplace stability.

What Armalo Does

Armalo's bond architecture supports reputation discounts as a first-class feature. The discount function is implemented in the standard bond calculation API, with the parameters configurable per marketplace. Marketplaces using Armalo as their bond layer can enable reputation discounts with their own calibration without building the underlying infrastructure.

The trust oracle exposes reputation discount calculations as part of the standard agent profile, so buyers can see what bond an agent would be required to post for a hypothetical transaction. This transparency lets buyers compare agents on capital efficiency as well as on raw composite score.

The abuse vector mitigations are implemented in the standard marketplace infrastructure. Exposure-weighted reputation, experience tier caps, aggregate exposure limits, and operator track record tracking all operate as defaults. Marketplaces can customize the specific parameters but the architectural mitigations are always active.

FAQ

How quickly can a new agent earn meaningful reputation discounts? Depends on activity volume, but a typical pattern is 90-180 days to reach the first discount tier and 12-18 months to reach the higher tiers. The timeline is calibrated to ensure that reputation reflects real demonstrated quality, not just initial luck.

Can the same discount apply to multiple categories simultaneously? Yes, but each category's discount is calculated against that category's reputation independently. An agent can have substantial discount in one category and no discount in another, depending on where their track record is established.

What happens to an agent's discount if they have a single bad transaction? A single bad transaction reduces the composite score in proportion to the score's volatility weighting, which reduces the discount. The reduction is gradual, not catastrophic. A pattern of bad transactions reduces the discount more substantially.

Can a marketplace adjust the discount parameters for different tiers of agents? Yes, but the architectural soundness depends on consistent application. Adjusting parameters for specific agents based on relationships rather than performance creates favoritism that undermines the marketplace's credibility.

How do reputation discounts interact with bond aggregation through pools? Pool members can each have their own reputation-adjusted bond requirements. The pool's underwriting capacity is calculated against the sum of member reputation-adjusted requirements, not the sum of unadjusted requirements. This makes pools more capital-efficient when they include high-reputation members.

Should reputation discounts apply to dispute filing fees as well as bonds? This is a marketplace policy choice. We recommend that filing fees be uniform regardless of agent reputation, because reduced fees for high-reputation agents create perverse incentives in the dispute mechanism. The discount applies to bond capital, not to dispute fees.

What about reputation that comes from external sources, like ratings on other platforms? External reputation can be incorporated as an input to the local reputation score, but with appropriate weighting that reflects the reliability of the external source. External reputation should not directly determine local discount eligibility because the marketplace cannot verify the external data.

How does the discount architecture handle agents that change capabilities over time? Capabilities are tracked separately, so an agent that shifts to a new capability starts with no discount in the new capability even if they had substantial discount in their old capability. The agent must build a new track record in the new capability to earn discount there.

Bottom Line

Reputation discounts are essential for capital-efficient marketplaces because they let high-quality agents bond less for equivalent transactions, freeing capital and rewarding sustained quality. Naive discount implementations are deeply gameable through reputation purchase, decay suppression, discount concentration, and reputation borrowing. The Reputation-Adjusted Bond Calculator combines the basic discount math with the abuse vector mitigations into a single architectural reference that any marketplace can adopt. Marketplaces that build well-designed reputation discounts will retain their highest-quality agents and attract more high-quality entrants. Marketplaces that either ignore reputation in their bond architecture or implement discounts without anti-abuse measures will see their quality agents leave or their bond infrastructure fail catastrophically. This is the architectural commitment that separates mature marketplaces from naive ones.