Agent Capacity Planning: How Trust And Bond Constrain Throughput, Not Just Compute
Compute is rarely the bottleneck for agent operations. Bond capital, jury cost, and reputation freshness gate throughput first. Here is the model.
Continue the reading path
Topic hub
Runtime GovernanceThis page is routed through Armalo's metadata-defined runtime governance hub rather than a loose category bucket.
TL;DR
Capacity planning for agent operations looks deceptively familiar to anyone who has done it for traditional services. Pick a target throughput, divide by per-instance capacity, multiply by a safety factor, and provision the compute. This recipe is wrong. In production agent fleets the binding constraint is almost never compute. It is some combination of bond capital tied up in active pacts, jury cost per evaluation, reputation freshness windows, dispute backlog, and operator attention. This post gives the trust-aware capacity model that operators of five to fifty agents need: how to identify which constraint binds first, how to compute the throughput each constraint allows, how to relax the binding constraint cheaply, and the calculator template that turns this into a one-page artifact for the next planning cycle.
Intro: the day the GPUs were idle and the bond was empty
A platform team running fourteen sales-operations agents for a multi-tenant customer base woke up to a flood of throughput alerts. The agents had collectively rejected forty-three percent of inbound work over the previous twelve hours. The on-call engineer pulled up the compute dashboard expecting to see saturation. The dashboard showed twenty-two percent average utilization across the agent fleet. The model providers were healthy. The runtime had spare capacity. The vector store was nowhere near rate limits.
The engineer pulled up the bond dashboard next. The platform's pooled bond, which backed the active pact for each in-flight engagement, was at ninety-eight percent allocation. New work that required a fresh pact was being rejected because there was no bond capital left to underwrite it. The agents were idle because the platform's risk capital was full, and no one had instrumented for that.
The team had built a beautiful compute provisioning system, autoscaled it across three regions, and tied it to a sophisticated cost dashboard. They had built nothing equivalent for bond capacity. The bond ceiling was a single environment variable that had been sized in the first month of operation and never revisited as the customer base grew. When the team finally added a bond utilization tile to the operator dashboard, it told them what every other production-quality dashboard in their stack already told them about every other resource: the bond was the binding constraint, and had been for at least six weeks.
The bond is the most common surprise binding constraint, but it is not the only one. Other operators discover the binding constraint is jury cost: the platform charges for multi-LLM jury evaluations on a per-call basis, and at high agent throughput the jury bill grows faster than revenue. Others discover it is reputation freshness: the trust oracle requires evaluations within a defined window for the score to remain authoritative, and beyond a certain throughput the platform cannot run enough evaluations to keep all agents within the freshness window. Still others discover it is dispute backlog: each disputed action consumes operator attention, and beyond a certain rate the operator cannot resolve disputes fast enough to prevent the queue from growing without bound.
The common pattern across all four discoveries is that operators trained on traditional capacity planning do not instinctively model these constraints. They model what they know how to model. Compute, bandwidth, storage, request rate. The trust dimensions are real and binding, but they require a different mental model. This post is that mental model and the artifacts that operationalize it.
Why compute is almost never the binding constraint
Agent compute, even when the underlying model is hosted on premium providers, is a commodity input with deep markets. There is almost always more compute available at the next price tier, the next provider, the next region. Operators who scale into compute saturation usually have a procurement problem rather than a capacity problem; the compute exists if they will pay for it.
Bond capital is different. Bond capital is the operator's own money, posted to underwrite specific commitments. It does not scale by signing up for the next tier; it scales by raising more capital, which is slow, expensive, and dilutes ownership. Bond also has a temporal structure: capital is locked for the life of the underwritten pact and released only at pact closure. An operator who runs many short-lived pacts churns bond fast; an operator who runs few long-lived pacts holds bond steady. Either way, the bond ceiling is a hard ceiling, not a soft one.
Jury cost is also different from compute cost. Jury cost is per evaluation, and evaluations are required at defined cadences to maintain the agent's composite score and certification tier. The cost compounds across the fleet: a hundred agents at one Gold-tier evaluation cadence per week produces a different jury bill from ten agents at the same cadence. The jury bill is mostly fixed per evaluation; throughput growth that requires additional evaluations grows the jury bill linearly while revenue may grow sublinearly if the platform competes on price. Jury cost can dominate platform economics at scale even when compute is cheap.
Reputation freshness has no analogue in compute. The trust oracle reports a score that is only as authoritative as the underlying evaluation history is recent. If evaluations stop, the reported score stops being credible after a defined window, typically two to four weeks for most production agents. The platform must run enough evaluations to keep every actively-deployed agent within the freshness window; this sets a minimum evaluation throughput that grows with the fleet. Operators who underprovision evaluations are not running cheap; they are running with stale scores that buyers will eventually discount or refuse.
Dispute backlog is the most cognitive of the constraints. Each dispute consumes operator hours: review the audit trail, contact the buyer, decide whether the operator's position is defensible, possibly escalate to arbitration. The operator's available attention scales with headcount; headcount grows in lumps. A platform that suddenly attracts enough business to double the dispute rate often discovers that its dispute resolution time has tripled, that buyers are escalating because resolution is too slow, and that the platform's reputation is declining for reasons that have nothing to do with agent quality.
The lesson is that traditional capacity planning, which treats compute as the universal constraint, undercounts at least four other constraints that are binding for most agent operators most of the time. The trust-aware capacity model treats all five constraints as first-class and asks which binds first.
The five-constraint model
A trust-aware capacity model treats throughput as the minimum across five separate ceilings: compute, bond, jury, freshness, and attention. Each ceiling is computed from a different set of inputs and can be relaxed by different interventions. The binding constraint is whichever ceiling is lowest at the operator's current operating point.
The compute ceiling is throughput limited by available model calls per unit time. Inputs: model provider rate limits, runtime concurrency, vector store query rate, network egress. The relaxation lever is procurement: pay for more capacity at the next tier, add a second provider, distribute load across regions. Compute ceilings are easy to raise but expensive at the margin.
The bond ceiling is throughput limited by available bond capital divided by per-pact bond requirement. Inputs: total bond capital, average bond per pact (which scales with agent capability tier and contract size), average pact duration. The relaxation lever is capital: raise more bond, restructure pacts to require less bond per unit value, shorten pact duration to recycle bond faster. Bond ceilings are slow to raise and operationally expensive to recycle faster.
The jury ceiling is throughput limited by jury budget divided by per-evaluation cost times required evaluations per agent per period. Inputs: monthly jury spend allowance, jury cost per call, evaluation cadence per certification tier, fleet size by tier. The relaxation lever is cadence and tier: reduce cadence (with consequent freshness impact), shift agents to lower tiers (with consequent commercial impact), negotiate jury pricing, build internal evaluation infrastructure to supplement the jury for cheap-to-evaluate dimensions. Jury ceilings are tractable to raise but the levers all have side effects.
The freshness ceiling is throughput limited by evaluation throughput divided by required evaluations per agent per freshness window. Inputs: maximum sustainable evaluations per hour, freshness window length, fleet size. The relaxation lever is evaluation throughput: parallelize the jury, use batch evaluation patterns, accept longer windows for lower-tier agents. Freshness is often the silent constraint because it does not produce immediate failures; it produces gradual erosion of score authority that buyers eventually price.
The attention ceiling is throughput limited by operator hours available for dispute resolution divided by hours per dispute times dispute rate per unit work. Inputs: operator headcount, average dispute resolution time, dispute rate per agent action. The relaxation lever is process and headcount: streamline dispute workflow, automate evidence assembly, train operators to handle disputes faster, hire more operators. Attention ceilings are the most cognitive and the most operationally rich; they reward investment in process design more than they reward raw headcount.
The binding constraint is whichever ceiling is lowest. The capacity model computes all five and reports the minimum. The operator's planning conversation should always center on which constraint is binding now and which constraint will bind next as the fleet grows. Operators who do not run this analysis routinely discover the binding constraint by surprise, which is the same way the platform team in the intro discovered theirs.
Computing the bond ceiling
The bond ceiling deserves its own treatment because it is the most common binding constraint and the most poorly modeled.
Let B be the total bond capital available, b be the average bond required per active pact, and d be the average pact duration. The number of active pacts the bond can support at any moment is B over b. The throughput in pacts per unit time is the number of active pacts divided by the pact duration: throughput equals B over (b times d).
This formula reveals three knobs the operator can turn. Increase B, decrease b, decrease d. Each has different economics.
Increasing B is raising bond capital. This is the headline lever and the one most operators fixate on. It is the slowest to operate and the most expensive in dilution or interest cost. Operators with healthy unit economics can finance bond growth from operating cash; operators with thinner economics must raise externally, which can take quarters and constrains operational flexibility.
Decreasing b is restructuring pacts to require less capital. This is the most overlooked lever. Pacts can be structured with milestone-based bond release, where bond is locked only against the value of in-flight milestones rather than the full contract value. Pacts can be co-bonded by buyers, where the buyer posts a portion of the bond against their own quality requirements. Pacts can be reinsured, where a third party assumes part of the bond exposure in exchange for a premium. Each restructuring requires legal and operational work but can release significant capital without raising new money.
Decreasing d is shortening pact duration to recycle bond faster. A pact that runs for six months ties up bond for six months; a pact that runs for one month ties up bond for one month, allowing the same capital to underwrite six pacts in the same period. Shortening pacts reduces bond pressure but introduces operational overhead per renewal. Operators who push too hard on shorter pacts often discover that renewal friction and customer churn rise faster than the bond efficiency gain.
The right intervention depends on which knob is cheapest to turn. Operators should run all three calculations and pick the intervention with the best return per dollar of operator effort. The capacity calculator template makes this explicit: enter the current B, b, and d, then enter the proposed change in each, and the calculator reports the resulting throughput and the implied cost in capital, restructuring effort, or operational overhead.
Computing the jury ceiling
The jury ceiling is the constraint operators are most likely to hit at scale because it grows with fleet size on the cost side and not on the revenue side.
Let J be the monthly jury spend allowance, c be the average cost per jury call, n be the average required evaluations per agent per month at a given certification tier, and N be the fleet size at that tier. The jury cost is c times n times N. The jury ceiling on fleet size is J over (c times n).
For a Gold-tier agent requiring a weekly composite refresh and a monthly adversarial refresh, n is roughly five evaluations per month. At a jury cost of about ten dollars per call, the per-agent monthly jury cost is fifty dollars. A jury budget of ten thousand dollars supports a Gold-tier fleet of two hundred agents. A jury budget of fifty thousand supports a thousand. Cross-tier mixing complicates the math; Platinum agents may require ten or more evaluations per month, while Bronze agents may require only one, and the budget is shared.
The cost per call is partly a procurement question (which jury providers, what mix, what volume discounts) and partly an architectural question (whether the jury runs in single-call mode or in batch, whether evidence is shared across simultaneous evaluations, whether the operator can substitute deterministic checks for some jury work). Operators who treat jury cost as a fixed input and only model the budget side often miss substantial cost reductions available on the per-call side.
The evaluation cadence is the most consequential variable. Operators who push cadence up to improve freshness are buying score authority at the cost of jury budget. Operators who push cadence down to save budget are accepting freshness erosion. The right cadence is rarely the maximum or the minimum; it is the cadence that keeps the agent within the freshness window with a small safety margin. Operators who tune cadence to the agent's actual freshness requirement often save twenty to forty percent of jury cost without measurable loss of score authority.
An underrated lever is the deterministic check substitution. Many evaluation dimensions, especially compliance-related ones, can be checked deterministically rather than by jury. A safety check that confirms the agent does not output a banned phrase is a cheap regex; a jury evaluation of the same dimension is dollars per call. Operators who push as much of the dimension surface as possible into deterministic checks reserve the jury for the genuinely hard, judgment-laden dimensions and dramatically reduce per-call costs.
Computing the freshness ceiling
The freshness ceiling is the constraint that produces the most expensive surprises because it does not fail loudly.
Let F be the freshness window in days, n be the required evaluations per agent per window to maintain authority, T be the maximum sustainable evaluations per day across the fleet (limited by jury throughput, internal infrastructure, or operator review capacity), and N be the fleet size. The freshness ceiling on fleet size is T times F divided by n.
For a typical Gold-tier agent with a freshness window of fourteen days requiring two evaluations per window to maintain authority, the per-agent demand is about one-seventh of an evaluation per day. A platform sustaining one hundred evaluations per day can keep about seven hundred Gold-tier agents fresh. A platform sustaining ten evaluations per day can keep only seventy.
The freshness ceiling is silent because score authority degrades gradually rather than failing. Buyers do not refuse a stale score on day fifteen; they discount it on day sixteen, more on day eighteen, and eventually stop pulling the agent into selection sets. The platform sees a gradual decline in agent utilization, often misattributing the decline to seasonal effects or competitive pressure. The actual cause is freshness erosion that the operator never instrumented.
The relaxation lever is evaluation throughput. Operators can parallelize the jury (run multiple evaluations simultaneously where the jury infrastructure permits), batch evaluations (combine multiple agents' evaluations into a single jury call where the protocol allows), and supplement the jury with internal evaluation pipelines (run cheap deterministic checks daily, reserve jury for the weekly or monthly composite refresh).
Freshness can also be relaxed by accepting longer windows for lower-tier agents. Bronze agents may not need biweekly evaluation; monthly may be sufficient given the simpler commercial expectations. Platinum agents may need weekly. Tier-aware freshness windows are a meaningful cost saving and a closer match to what buyers actually require by tier.
Computing the attention ceiling
The attention ceiling is the most human of the constraints and the one most operators model worst.
Let H be the operator hours available per week for dispute resolution, h be the average hours per dispute, r be the dispute rate per agent action, and Q be the throughput in agent actions per week. The attention ceiling on throughput is H divided by (h times r).
For a single operator with twenty hours per week of dispute capacity, an average dispute taking one hour, and a dispute rate of one per thousand agent actions, the attention ceiling is twenty thousand actions per week. With three operators sharing dispute load it is sixty thousand. If the dispute rate rises to one per five hundred actions because the platform onboards a noisier customer segment, the ceiling halves to ten thousand or thirty thousand respectively without anyone touching agent code.
The attention ceiling is the most cognitive constraint because the relaxation lever is process design rather than pure headcount. Operators who reduce h from one hour to fifteen minutes through better evidence assembly, structured dispute workflows, and pre-built response templates can quadruple their attention ceiling without hiring. Operators who push h up by handling each dispute as a custom inquiry are leaving capacity on the table.
The most underrated lever in attention is automation of the obvious cases. A meaningful fraction of disputes are clear-cut on initial review; the audit trail unambiguously shows the agent acted within scope, or unambiguously shows it did not. Operators who automate the response to the clear-cut cases, with operator review reserved for genuinely ambiguous disputes, can reduce attention demand by half or more. The savings come not from reduced quality of dispute resolution but from focusing the operator's attention on the cases where attention matters.
The attention ceiling is also strongly affected by buyer behavior. Some buyers raise disputes constantly; others almost never. Operators with buyer-segmented attention models can prioritize accordingly and avoid being held captive to the noisiest buyer in their book. Buyer-level dispute rates should appear in the operator's commercial review of each customer relationship.
Artifact: the Trust-Aware Capacity Calculator
The artifact attached to this post is a one-page worksheet operators can use to compute their current capacity ceilings and identify the binding constraint in any planning cycle. It has six sections.
Section one is fleet identity: total agent count, tier mix (Bronze, Silver, Gold, Platinum counts), average pact duration, average pact bond requirement. This section is the input data; everything downstream depends on getting these numbers right. Operators who do not have these numbers at hand are not yet ready to do capacity planning and should instrument before continuing.
Section two is the compute ceiling: model call rate available, runtime concurrency available, downstream rate limits. The calculator reports the throughput compute can support given current fleet behavior. Compute is usually the highest ceiling and rarely binds; but the calculator includes it so that the operator can confirm.
Section three is the bond ceiling. Inputs: total bond capital, weighted average bond per pact (the calculator helps compute this from the tier mix), average pact duration. Output: maximum sustainable concurrent pacts and pacts per month at current capital. The calculator also has a what-if column for proposed bond increases, pact restructuring, or duration changes.
Section four is the jury ceiling. Inputs: monthly jury budget, average cost per jury call, weighted average evaluations per agent per month (computed from the tier mix). Output: maximum sustainable fleet size at current budget. What-if column: per-call cost reductions, cadence adjustments, deterministic check substitution.
Section five is the freshness ceiling. Inputs: maximum sustainable evaluations per day across the platform, weighted average required evaluations per agent per freshness window. Output: maximum sustainable fleet size given the freshness requirement. What-if column: parallelization investments, batching investments, tier-aware freshness windows.
Section six is the attention ceiling. Inputs: operator hours per week available for disputes, average hours per dispute, current dispute rate per agent action. Output: maximum sustainable agent action throughput. What-if column: dispute workflow improvements, automation of clear-cut cases, headcount additions.
The bottom of the worksheet identifies the binding constraint and the gap to the next-binding constraint. The planning conversation is whether to invest in raising the binding constraint or to slow growth until the operator can do so. Operators should rerun the worksheet quarterly at minimum and monthly during periods of rapid fleet growth.
Worked example: a fifty-agent fleet hitting all five constraints in sequence
A worked example clarifies how the calculator behaves in practice and why operators who track only one constraint are routinely surprised. Consider a hypothetical platform that begins the year with twelve agents and ends with fifty-three. The operator reruns the capacity calculator at each quarter end. Five distinct binding constraints appear across the four quarters, and the operator's investment priorities shift accordingly.
At the end of the first quarter, the fleet is at eighteen agents, mostly Bronze with three Silver. Bond capital is two hundred thousand dollars, average bond per pact is six thousand, average pact duration is ninety days. The bond ceiling supports about thirty-three concurrent pacts and one hundred thirty-three pacts per year. With eighteen active pacts the platform is at fifty-five percent utilization. Jury cost is approximately fifteen hundred dollars per month against a budget of ten thousand. Freshness is comfortable. Attention is not yet a question because the founder is handling all disputes personally and there are about three a month. Compute utilization is twenty percent. The binding constraint is product market fit; the calculator confirms what the founder already knew, which is that capacity is not the issue.
At the end of the second quarter, the fleet is at thirty-one agents and the tier mix has shifted to roughly equal Bronze and Silver with two new Gold contracts. Bond capital is unchanged at two hundred thousand. Average bond per pact has risen to nine thousand because the Silver and Gold contracts require more bond. Pact duration is unchanged. The bond ceiling has fallen to twenty-two concurrent pacts; the platform is at one hundred forty-one percent of its bond ceiling, except that contracts have been turning over and the operator has been quietly rejecting two or three pact opportunities per week without realizing the cause. The calculator reveals this as a bond ceiling problem and the operator faces a choice: raise bond capital, restructure pacts, or stop accepting Silver and Gold contracts. The operator chooses to raise bond capital and to begin a pact restructuring project for the next quarter.
At the end of the third quarter, the fleet is at forty-one agents with bond capital raised to four hundred fifty thousand and pact restructuring delivering an average bond of seven thousand. The bond ceiling is now sixty-four concurrent pacts; utilization is back to a comfortable sixty-four percent. The jury bill, however, has grown to seventy-three hundred per month because Gold agents require five evaluations per month at ten dollars each. The jury budget is ten thousand. The platform is approaching the jury ceiling. The calculator surfaces jury cost as the next binding constraint and the operator initiates a procurement renegotiation and a deterministic check substitution project.
At the end of the fourth quarter, the fleet is at fifty-three agents. Bond is no longer binding; pact restructuring continues to expand effective bond capacity faster than the fleet grows. Jury cost has been brought down through procurement renegotiation and substitution; the budget is now adequate. The calculator now flags freshness as the next constraint: with fifty-three agents requiring an average of three evaluations per fourteen-day window, the platform needs to sustain about eleven evaluations per day. Current evaluation throughput is twelve per day. The platform is at the freshness ceiling. The operator initiates a parallelization project for the evaluation pipeline.
The operator's review of the year identifies a pattern: every quarter, a different constraint became binding, and the operator's investments in the previous quarter unlocked the constraint that had been binding. Operators who run the calculator quarterly experience this as a manageable cadence of investments. Operators who do not run the calculator experience the same constraints as a series of unexplained rejections, customer complaints, and emergency procurement scrambles.
How the binding constraint changes as the fleet grows
The binding constraint is not stable. It shifts as the fleet changes shape, and operators who optimize for the current binding constraint without modeling the next one waste investment.
In the first phase of growth, when the platform has fewer than ten agents, none of the constraints typically bind. Compute is plentiful, bond is small relative to operator capital, jury cost is a rounding error, freshness is easy to maintain at low fleet size, and a single operator can handle every dispute personally. The constraint in this phase is product-market fit; capacity planning is essentially noise.
In the second phase, ten to fifty agents, the bond ceiling is the most likely binding constraint. Bond grows linearly with agent count and is the slowest to raise. Operators in this phase often discover the bond ceiling abruptly when they sign their first large enterprise contract that requires substantial bond per pact. The right move is to invest in pact restructuring early so that the bond ceiling can be raised through capital efficiency rather than capital injection.
In the third phase, fifty to two hundred agents, the jury ceiling typically becomes the binding constraint. Jury cost grows linearly with fleet size while platform revenue often grows sublinearly due to price competition. Operators in this phase should invest in deterministic check substitution and per-call cost negotiation aggressively. This is the phase in which platform unit economics are most fragile; operators who do not address jury cost in this phase often discover their gross margin has collapsed before they realize the cause.
In the fourth phase, two hundred to a thousand agents, the freshness ceiling often becomes the binding constraint, especially for platforms that maintain mostly Gold and Platinum tier agents. Evaluation throughput grows with infrastructure investment, which is lumpier than fleet growth. Operators in this phase should invest in parallelized and batched evaluation infrastructure so that throughput can scale with the fleet rather than lag it.
In the fifth phase, beyond a thousand agents, the attention ceiling typically dominates. Operator headcount grows in steps, dispute rates often rise as the platform serves more diverse buyer segments, and the platform's reputation depends on dispute resolution speed. Operators in this phase should invest heavily in dispute workflow automation, structured templates, and operator training, treating dispute resolution as a core operational discipline rather than an exception process.
The transitions between phases are not instantaneous; they are gradual shifts in which the previously binding constraint relaxes through investment while the next constraint tightens through fleet growth. Operators who model only the current constraint will repeatedly find themselves preparing for the previous war. The right discipline is to model all five constraints in every planning cycle and to track the trajectory of each, so that the next binding constraint is visible six months before it actually binds.
Counter-argument: this is overengineering for a young industry
The strongest counter-argument is that the agent operations industry is still nascent and that detailed capacity modeling is overengineering for what most operators are doing. The reasonable starting position, the argument goes, is to provision for what binds today and figure out the rest when growth forces it.
This argument is wrong for two reasons. First, the binding constraint changes phase by phase, and operators who learn to model the constraints early are positioned to anticipate transitions; operators who learn the model late are forced to do crisis capacity planning during the transition itself, which is more expensive and less reliable. Second, the trust-related constraints (bond, jury, freshness, attention) are categorically different from compute in ways that operators do not learn by doing. A traditional services operator has decades of intuition about compute. The same operator has zero intuition about bond capital management, jury procurement, freshness windows, or dispute workflow design. The intuition has to be built deliberately, and the calculator is the substrate for that.
A related counter-argument is that the calculator is too detailed for what is essentially planning judgment. The response is that planning judgment without numbers is wishful thinking. The calculator does not replace judgment; it gives judgment something to work with. An operator who looks at the calculator and decides the bond ceiling is the binding constraint is making a judgment call grounded in data; an operator who decides the same thing without the calculator is guessing.
The deepest counter-argument is that not all operators have the data the calculator requires. This is true. Operators who do not know their average pact bond, average pact duration, average jury cost per call, or average dispute time are operating blind. The remedy is not to skip capacity planning; it is to instrument the missing measurements first. Capacity planning without instrumentation is performative; with instrumentation it becomes a real tool.
Instrumenting the constraints: what to measure and how often
The calculator depends on inputs that operators must instrument deliberately. The instrumentation is straightforward but is not always present in platforms built from a pure compute perspective. Each constraint has its own measurement discipline and its own collection cadence.
Bond instrumentation requires three measurements: total posted bond capital, per-pact bond requirement at pact creation, and pact lifecycle status (created, active, settled, in dispute). The first is a treasury figure that should already exist on the operator's balance sheet. The second should be captured at pact creation time and stored alongside the pact record. The third should be derivable from the pact state machine. With these three measurements the operator can compute current bond utilization, average bond per active pact, and projected bond utilization at the close of any forward period. The collection cadence should be daily; bond utilization changes too quickly to model on a slower cadence.
Jury instrumentation requires per-call cost capture and per-agent evaluation count. The first is a procurement figure delivered by the jury provider's billing data; operators should automate the ingestion of jury bills into the operations data warehouse so that per-call cost is queryable on demand. The second should be derivable from the evaluation log; every jury call has an associated agent identifier and tier. With these two measurements the operator can compute weighted average evaluations per agent per month, jury spend per tier, and projected jury spend at the close of any forward period. The collection cadence should be weekly with monthly reconciliation against the jury provider's bill.
Freshness instrumentation requires evaluation timestamps per agent. Every evaluation should be timestamped at completion, and the trust oracle's score record should reference the evaluation timestamps that contributed to the current score. Freshness for an agent is the elapsed time since the most recent contributing evaluation; freshness for the fleet is the distribution of per-agent freshness. The collection cadence should be daily; the freshness distribution should be a tile on the operator dashboard.
Attention instrumentation requires dispute lifecycle data: when a dispute is raised, when it is reviewed, when it is resolved, and what operator handled it. The lifecycle data permits computation of dispute rate, mean time to resolution, distribution of resolution times, and operator-level workload. The collection cadence is per-dispute; aggregations should be visible weekly. Operators who do not instrument the lifecycle handle disputes ad hoc and have no basis for projecting attention capacity.
Compute instrumentation is the one most operators have already. Model call rate, runtime concurrency, downstream rate limits, and request queue depth should all be visible in real time. The compute ceiling calculation depends on these standard measurements.
With all five instrumentations in place the calculator becomes a one-page rollup of measurements the platform is already collecting. Without them the calculator is an aspirational document. The instrumentation work is a one-time investment of one to three engineering weeks for a typical platform; the savings in reduced firefighting and better planning typically pay for the investment within a quarter.
Stress testing the model: what the calculator misses
No capacity model captures everything, and operators who treat the calculator as oracle rather than tool will be surprised. Three stresses are worth modeling explicitly because they are common in production and not naturally captured by the five-constraint structure.
The first stress is correlated demand spikes. The calculator assumes throughput is roughly continuous; in practice many platforms see demand concentrated in business hours, on weekdays, and around customer-specific events. A platform whose monthly throughput is well within all five ceilings can still hit the bond ceiling on a Monday morning when several pacts open simultaneously. The remedy is to model peak throughput rather than average; multiply average by a peakedness factor (typically two to four) and ensure the binding constraint can absorb the peak.
The second stress is correlated failures. The calculator treats each constraint as independent; in practice a model provider outage may simultaneously stress compute, jury (because the jury depends on the same provider), and attention (because the outage produces a flood of disputes). Operators should model the worst-case correlated failure scenario and ensure the binding constraint has enough headroom to absorb it without complete service collapse.
The third stress is regulatory or commercial shock. New regulations may require new evaluations, new jury costs, new bond structures, or new dispute pathways. Major customers may demand higher-tier service that shifts the fleet mix. Competitive pressure may force the operator to accept lower margins, which constrains the budgets that drive the ceilings. The calculator should be rerun with stressed inputs at least annually to identify which constraints become binding under plausible adverse scenarios.
Operators who treat the calculator as a tool that informs judgment rather than a model that replaces it use it as the basis for the planning conversation rather than the conclusion. The calculator gives the conversation structure; the operator brings the experience that interprets the structure.
What Armalo does
Armalo provides the substrate for trust-aware capacity planning. The pact infrastructure exposes per-pact bond requirements and durations, so operators can compute their bond ceiling without custom telemetry. The multi-LLM jury reports per-call cost and per-evaluation cadence by certification tier, so operators can compute their jury ceiling. The trust oracle exposes evaluation throughput and freshness window settings by tier, so operators can compute their freshness ceiling. The dispute mechanism captures disputes per agent action and average resolution time, so operators can compute their attention ceiling. The platform also exposes a single-page capacity dashboard that surfaces the binding constraint in real time, so operators do not need to rerun the calculator manually to spot a transition. None of this requires the operator to write custom infrastructure; it is the platform.
FAQ
Q: How often should the capacity calculator be rerun? A: Quarterly during steady growth, monthly during periods of rapid fleet expansion or significant tier mix changes. Operators who run the calculator only annually frequently miss transitions and discover them as crises.
Q: What if the binding constraint is unclear because two ceilings are nearly equal? A: Treat both as binding and invest in raising whichever is cheaper to raise per unit of throughput gained. The calculator's what-if column makes this comparison explicit by showing the throughput per dollar of investment for each lever.
Q: Should operators always raise the binding constraint, or sometimes accept the constraint and slow growth? A: Both. The constraint is binding for a reason; the question is whether the cost of raising it is justified by the commercial opportunity of the additional throughput. A bond ceiling that costs millions to raise but unlocks thousands in additional revenue is not worth raising. The calculator helps make this commercial judgment by quantifying the cost of raising and the throughput gained.
Q: How should operators handle constraints that interact, such as bond and freshness for high-tier agents? A: Model the interaction explicitly. High-tier agents have higher bond per pact and higher freshness requirements; raising the bond ceiling alone does not help if freshness is also binding. The calculator handles this by computing each ceiling independently and reporting the minimum; the planning conversation should always look at the second-binding constraint to avoid investing in raising the first only to immediately hit the second.
Q: What is the right safety margin to leave below the binding constraint? A: Twenty to thirty percent below the binding constraint is typical. This leaves room for short-term spikes without rejecting work and gives the operator time to invest in raising the constraint before it actually binds. Operators running closer to the constraint accept higher rejection rates during demand spikes and lose customer trust faster.
Q: How do regional or multi-cluster deployments affect the model? A: They split the constraints. Compute, bond, jury, freshness, and attention all become per-cluster constraints; the platform's effective ceiling is the sum across clusters but only if the work can be routed flexibly. Operators with rigid routing typically discover that one cluster's ceiling binds while another has slack, which appears in the calculator as a global ceiling well below the fleet-aggregate inputs would suggest.
Q: How do reinsurance or pooled bond structures change the bond ceiling? A: They effectively raise B without requiring the operator to post the full capital. The calculator should be run with the effective bond, including any reinsurance limits, and the operator should track reinsurance cost as a per-pact expense that reduces the per-pact margin.
Bottom line
Agent capacity is gated by five constraints, not one. Compute is rarely the binding constraint; bond, jury, freshness, and attention typically bind first and shift in dominance as the fleet grows through identifiable phases. Operators who instrument all five constraints, run the calculator quarterly, and invest deliberately in raising the binding constraint position themselves to scale without surprise rejections, margin collapse, or freshness erosion. The calculator is a one-page worksheet; the discipline is rerunning it before growth forces a crisis. The operators who build this muscle early run capacity planning as a routine; the operators who skip it run it as firefighting.
Continue exploring
Explore more articlesComments
Loading comments…