Escrow For Long-Horizon Tasks: When The Job Lasts Six Weeks And The Bond Lasts Sixty

TL;DR

Most agent escrow infrastructure is designed for short-horizon transactions where the work completes in hours and any damage surfaces within days. This works fine for the bulk of current agent activity. It breaks down completely for the long-horizon work that the agent economy is moving toward — multi-week or multi-month engagements where the agent does substantive work over time and the buyer's exposure to defective output extends well past the formal completion date. For these engagements the standard bond architecture is inadequate. The bond needs to be staged across pre-engagement, in-flight, post-completion, and dispute-window phases, with different sizing and release rules at each stage. This post lays out the Long-Horizon Bond Schedule as a structured framework for designing these escrow arrangements, walks through the math at each stage, and provides a worked template you can adapt for your own engagements. The marketplaces that solve this problem will unlock a class of transactions that are currently uneconomic to bond.

Intro: The Six-Week Job And The Sixty-Day Damage Window

In early 2026 a mid-market software firm engaged an autonomous agent for a six-week database migration project. The agent was reputable, the pact was carefully written, and the marketplace required a bond proportional to the project value. The engagement completed on schedule. The agent was paid, the bond was released, and everyone moved on.

Forty-three days later the migration was discovered to have introduced a subtle data corruption issue that was caught only when an unrelated quarterly audit reconciled aggregate revenue figures against detailed transaction records. The corruption traced back to a malformed data type conversion in one of the migration scripts the agent had produced. The remediation cost was about $74,000 — restoration of corrupted records from backups, manual reconciliation of the quarterly figures, accountant time, and a delayed quarterly close that affected investor relations.

The agent was no longer accessible because the company had switched to other agents. The bond had been released six weeks earlier. The marketplace had no remaining mechanism to compensate the buyer because the dispute window had closed. The buyer absorbed the $74,000.

When the buyer raised the issue with the marketplace, the operator's response was that the bond mechanism was working as designed. Bonds were held during the project and released on completion. If the buyer wanted protection against latent defects, they should have negotiated for an extended bond holdback or warranty escrow. The buyer's counter-argument was that the marketplace's standard bond was implicitly representing protection that it did not actually provide for long-horizon work, because the timing of the bond release did not match the timing of the realistic damage discovery for that class of work. The marketplace and the buyer ended up in a multi-month dispute that exhausted both parties without compensating anyone.

This case illustrates the core problem with applying short-horizon bond architecture to long-horizon work. For a one-hour task, the bond can be released within hours of completion because most failures surface within hours. For a six-week project, the bond probably needs to be partially held for sixty or ninety days post-completion because that is the realistic window in which latent defects surface. The standard bond architecture does not support this kind of staged commitment, and the result is that long-horizon work is structurally under-protected even when it is nominally bonded.

This post is the design document for a Long-Horizon Bond Schedule that handles the staging properly. It addresses pre-engagement bonding to ensure agents are committed before work starts, in-flight bonding to ensure agents continue to perform during long projects, post-completion bonding to cover the latent defect window, and dispute-window bonding to handle claims filed during the formal protection period. By the end of the post you should be able to design escrow arrangements for any long-horizon engagement and know what bond sizing makes sense at each stage.

Section One: Why Short-Horizon Architecture Fails Long-Horizon Work

The short-horizon bond architecture has three structural assumptions that fail for long-horizon work.

The first assumption is that the bond duration corresponds to the work duration. For a one-hour task, the bond is locked for one hour plus a brief dispute window. For a one-day task, the bond is locked for one day plus the dispute window. The assumption is that the bond's economic exposure period scales linearly with the work period.

This is approximately right for short-horizon work because the failure modes are typically observable during or shortly after the work completes. Code that crashes does so on the first run. Communication that misfires does so when sent. Data that is malformed becomes visible quickly. The damage discovery window is short and the bond duration matches.

For long-horizon work, the failure modes have a different structure. Many failures are subtle and surface only when the buyer's downstream systems exercise the work in ways that were not anticipated during the engagement. A migration script may corrupt one record in a thousand, surfacing only when an audit catches the inconsistency. A research recommendation may turn out to be wrong only when the buyer acts on it and observes the consequences months later. An infrastructure change may degrade performance subtly, surfacing only as cumulative cost over time. The damage discovery window for long-horizon work is often much longer than the work duration itself, and the bond duration based on work duration is therefore inadequate.

The second structural assumption is that the bond is a single static commitment that exists from start to finish of the engagement. The agent posts the bond when the pact is created, the bond sits there during the work, and it is released when the work completes. There is no mechanism to adjust the bond level during the work, to add additional bond at sensitive stages, or to extend the bond past completion at variable amounts.

For short-horizon work this is adequate because the work proceeds quickly and the bond can be sized for the entire engagement at the outset. For long-horizon work the bond needs to be more like a staged commitment with different amounts at different phases. The work has natural inflection points where bond increase or decrease makes sense, and the standard architecture cannot support this.

The third structural assumption is that the dispute window is a fixed duration after completion that applies uniformly to all categories. Most marketplaces use a dispute window of 7-30 days regardless of the work category. For short-horizon work this is fine because the discovery window is also short. For long-horizon work the dispute window may need to be 60, 90, or 180 days depending on the category and the buyer's exposure profile. A 14-day dispute window on a six-week migration project is structurally inadequate for the realistic discovery timeline.

The Long-Horizon Bond Schedule addresses all three of these assumptions by replacing the single static bond with a four-stage staged bond, replacing the fixed dispute window with a category-calibrated extended window, and adding mechanisms for bond adjustment during the engagement.

Section Two: The Four Stages Of Long-Horizon Bonding

The Long-Horizon Bond Schedule divides the engagement lifecycle into four stages, each with its own bond sizing, release rules, and risk profile.

Stage one is pre-engagement bonding, covering the period from pact creation to project kickoff. During this stage the agent has committed to the engagement but has not yet started substantive work. The bond purpose at this stage is to ensure the agent will actually show up and do the work. The risk being underwritten is no-show or early withdrawal.

The pre-engagement bond should be sized at roughly 5-10% of the project value, with the lower end appropriate for established agents and the higher end appropriate for newer agents without track record. The bond is released to the buyer if the agent fails to start work within the agreed kickoff window. The bond is rolled forward to stage two if the agent successfully starts the engagement.

Stage two is in-flight bonding, covering the period from project kickoff to substantive completion. During this stage the agent is actively producing work and the buyer is reviewing it. The bond purpose is to ensure the agent continues to perform and does not abandon the engagement partway through. The risk being underwritten is mid-engagement default or substandard performance.

The in-flight bond should be sized at 15-25% of the remaining project value, with adjustments at major milestones. As each milestone completes and the buyer accepts the deliverable, a portion of the in-flight bond is released to the agent and the remaining bond is rebalanced against the remaining work. This staged release provides the agent with intermediate liquidity and reduces the bond capital tied up over the long engagement.

Stage three is post-completion bonding, covering the period from substantive completion to the end of the latent defect window. During this stage the agent's work is complete and the buyer is exercising it in production. The bond purpose is to cover compensation for defects discovered after completion. The risk being underwritten is latent defect liability.

The post-completion bond is the critical stage that most current architectures handle poorly. The bond should be sized at 25-40% of the project value, depending on the category. For low-risk categories the lower end is appropriate; for high-risk categories like infrastructure and financial work the higher end is appropriate. The bond is held for the full latent defect window without partial release.

Stage four is dispute-window bonding, covering the period from the end of the latent defect window to the close of all open disputes. During this stage no new claims can be filed but existing disputes are still being resolved. The bond purpose is to ensure that any claims filed during the latent defect window can be paid out even if resolution takes additional time. The risk being underwritten is dispute resolution latency.

The dispute-window bond should equal the maximum claim amount across any open disputes from the latent defect window. If no disputes are open, the dispute-window bond is zero and the post-completion bond is fully released. If disputes are open, the dispute-window bond is held until all are resolved, after which any remaining amount is released to the agent.

Taken together, the four stages provide continuous bond coverage from pact creation through final dispute resolution. The total bond value tied up at any moment is much smaller than the sum of all four stages because the bond rolls forward from stage to stage rather than being posted independently at each stage. This makes the architecture capital-efficient while still providing meaningful protection across the full engagement lifecycle.

Section Three: The Latent Defect Window By Category

The duration of the post-completion bond stage depends on the category of work, because different categories have different empirical timelines for latent defect discovery. Calibrating the latent defect window correctly is the key to making post-completion bonding work without tying up bond capital indefinitely.

We have collected empirical data on latent defect discovery timing across nine categories. The 95th percentile discovery time varies from 7 days to 180 days depending on the category.

Text generation for low-stakes content has a latent defect window of about 7 days. Most issues with marketing copy or documentation are caught quickly because the work is exercised by reading it.

Data extraction and structuring has a latent defect window of about 21 days. Issues with extracted data often surface when the data flows downstream into systems that exercise it in different ways.

Non-production code has a latent defect window of about 14 days. Prototype code and exploratory analysis are usually exercised quickly enough that issues surface within two weeks.

Production code has a latent defect window of about 60 days. Issues with production code often surface only when production load patterns differ from test patterns, when seasonal variation exposes edge cases, or when integration with other systems triggers unexpected interactions.

Autonomous communication has a latent defect window of about 14 days. Most communication issues surface quickly through customer reactions or compliance reviews.

Autonomous transactions has a latent defect window of about 30 days. Transaction issues often surface during reconciliation cycles that operate on monthly or quarterly cadence.

Autonomous research has a latent defect window of about 90 days. Research recommendations may not be actioned for weeks or months after delivery, and the consequences of acting on bad research may take additional time to materialize.

Autonomous infrastructure has a latent defect window of about 90-180 days. Infrastructure issues can take months to surface because they often manifest as gradual degradation or as failures that occur only under specific load patterns that take time to recur.

Autonomous physical-world interaction has a latent defect window of about 60-180 days, with high variance based on the specific physical context.

These empirical baselines are starting points and should be adjusted based on the specific context of each engagement. A migration into a system with weekly audit cycles may have a shorter effective latent defect window than the category baseline because the audit catches issues quickly. A migration into a system with annual audit cycles may have a longer effective window. The pact should specify the effective latent defect window based on the buyer's actual operational profile rather than relying on the category baseline alone.

The practical implication is that long-horizon engagements often have post-completion bond stages that last longer than the work itself. A six-week migration into a system with quarterly audits may need a 90-day post-completion bond, meaning total bond coverage for the engagement runs about four months. This is a substantial commitment of bond capital and it is one of the reasons that bond aggregation through mutual pools (covered in an earlier post in this series) is essential for making long-horizon work economically viable.

Section Four: Bond Adjustment During The Engagement

A meaningful difference between long-horizon and short-horizon bonding is the need for bond adjustment during the engagement. As the work progresses, the appropriate bond level changes based on the work completed, the milestones achieved, and the risks exposed.

Three adjustment mechanisms are useful in long-horizon engagements.

The first adjustment is milestone-driven release. As each milestone completes and the buyer accepts the deliverable, a portion of the in-flight bond is released to the agent. This provides the agent with intermediate liquidity and reduces the capital tied up. The release amount should be calibrated so that the remaining bond still covers the remaining work, with the release proportion typically 60-75% of the milestone value.

The second adjustment is risk-driven increase. If the engagement enters a particularly sensitive phase — for example, the cutover phase of a database migration where the actual production data is being touched — the bond should increase to reflect the elevated risk during that phase. The increase comes from new bond posted by the agent or from a contingent bond reservation that was set up at engagement start. After the sensitive phase completes, the bond returns to the baseline level.

The third adjustment is scope-driven update. If the engagement scope changes — additional work added, requirements modified, deliverables expanded — the bond should be recalibrated against the updated scope. Scope changes are common in long-horizon engagements and the bond architecture needs to handle them gracefully rather than treating the original bond as fixed.

These adjustment mechanisms add complexity to the bond infrastructure but they are essential for long-horizon work. A static bond that does not respond to the actual progression of the engagement will either be over-bonded for most of the work (tying up capital unnecessarily) or under-bonded at sensitive phases (providing inadequate protection at the highest-risk moments). Dynamic adjustment provides the right bond level at each phase.

The administration of bond adjustment requires careful pact design. The pact should specify the adjustment triggers explicitly: the milestones that release bond, the events that increase bond, the scope changes that recalibrate bond. Adjustments that are not specified in the pact require renegotiation, which is friction. Adjustments that are specified can execute automatically as the relevant events occur.

Section Five: The Long-Horizon Bond Schedule Template

The artifact for this post is the Long-Horizon Bond Schedule Template, a structured worksheet that any buyer or agent can use to design the bond architecture for a specific engagement. The template has six sections.

Section one captures the engagement basics: project value, expected duration, capability category, milestone structure. This information drives the calculations in the subsequent sections.

Section two specifies the pre-engagement bond. The default sizing is 5-10% of project value, adjustable based on the agent's track record. The release condition is failure to kick off within the agreed window.

Section three specifies the in-flight bond. The default sizing is 15-25% of remaining project value, with milestone-driven release schedule. The template includes the milestone definitions and the release amounts at each milestone.

Section four specifies the post-completion bond. The default sizing is 25-40% of project value, with the latent defect window calibrated to the category. The template includes the discovery timeline assumption and the rationale for the chosen window length.

Section five specifies the dispute-window bond, which is contingent on disputes being open at the end of the latent defect window. The template includes the procedure for calculating the dispute-window bond based on open dispute amounts.

Section six specifies the bond adjustment triggers: milestones that release bond, events that increase bond, scope changes that recalibrate bond. The template includes a table of common triggers and the associated bond changes.

The template produces a complete bond schedule that can be embedded in the pact as a structured commitment. The schedule defines the bond level at each phase of the engagement, the events that trigger bond changes, and the release procedure at the close of the dispute window. Both buyer and agent can refer to the schedule throughout the engagement to know what bond level applies at any given moment.

We have published the full template as a downloadable document linked below. Buyers and agents are welcome to use it as a starting point for their own engagements. We have also included worked examples for three common long-horizon engagement types: a database migration, an autonomous research project, and an infrastructure operations contract. The examples show how the template applies in practice and how the bond sizing varies across categories.

Section Six: Edge Cases In Long-Horizon Bonding

Three edge cases deserve direct treatment because they come up regularly in long-horizon engagements.

The first edge case is partial completion. If an engagement terminates partway through — by mutual agreement, by buyer cancellation, or by agent withdrawal — the bond architecture needs to handle the partial state gracefully. The standard handling is to evaluate the work completed at the termination point, release the corresponding portion of the in-flight bond to the agent, and convert the remaining bond into a partial post-completion bond covering the work that was actually delivered. This preserves protection for the buyer on the partial work while not over-bonding the engagement that did not complete.

The second edge case is dispute during the in-flight stage. If the buyer files a dispute about work delivered at a milestone, the in-flight bond should not release until the dispute resolves. This is straightforward in principle but the operational handling requires pact-level specification of how mid-engagement disputes affect the bond release schedule. The pact should specify whether milestone releases are paused during disputes or whether they proceed with bond reservation against the disputed amount.

The third edge case is engagement extension. If the engagement runs longer than originally scoped, the bond schedule needs to extend correspondingly. The standard handling is to recalibrate all bond stages against the new completion timeline, with the agent posting additional bond to cover the extension. This requires the pact to specify the procedure for extension, including the bond adjustment mechanism.

None of these edge cases is particularly difficult to handle if the pact is well-designed. They become difficult only when the pact is silent on them, leaving the parties to negotiate in real time during the engagement. A pact that anticipates these cases at signing avoids the friction of mid-engagement renegotiation.

Section Seven: How Long-Horizon Bonding Changes The Marketplace

The availability of well-designed long-horizon bond architecture changes what kinds of work can flow through agent marketplaces. Three categories of work that are currently underserved become viable.

The first category is multi-week professional services engagements. These have historically required direct relationships between buyers and providers because marketplace bond architecture could not handle the engagement length. With long-horizon bonding, these engagements can flow through marketplaces with the same trust signals and dispute resolution that short-horizon transactions enjoy.

The second category is project-based work with substantial latent defect risk. Database migrations, infrastructure transitions, software ports, security audits — these all have failure modes that surface long after the work completes. Without post-completion bonding, marketplaces cannot offer credible protection for this kind of work. With it, they can.

The third category is ongoing service relationships. Long-running contracts where an agent provides services on a continuous basis, with the engagement extending for months or years, have historically required custom contract negotiation outside marketplace infrastructure. With staged bond architecture, marketplaces can support these relationships through ongoing bond updates that adjust to the evolving engagement.

The expansion of marketplace-supported work into these categories represents a substantial increase in the addressable market for agent marketplaces. The marketplaces that build long-horizon bond architecture early will capture the share of this expansion. The marketplaces that stick with short-horizon architecture will be limited to the current set of categories.

Counter-Argument: Long-Horizon Bonding Is Too Capital-Intensive

The most common objection to long-horizon bond architecture is that it ties up too much capital. A six-week engagement with a 90-day post-completion bond requires capital commitment for over four months, which is a substantial cost of capital relative to short-horizon transactions.

This objection is real but it conflates two distinct issues.

The first issue is whether long-horizon bonding is more capital-intensive than short-horizon bonding for equivalent total work. The answer is no. A six-week project bonded with a four-month staged bond uses less aggregate capital than the equivalent volume of work spread across many short-horizon transactions, each independently bonded. The staged bond is capital-efficient for what it covers.

The second issue is whether long-horizon bonding is more capital-intensive than the current practice of short-horizon bonding plus extra-marketplace warranty arrangements. For most current long-horizon work, buyers either accept the marketplace's inadequate post-completion protection (and absorb latent defect risk) or negotiate separate warranty escrows outside the marketplace (which carry their own capital cost). When you add up the total cost of capital for protected long-horizon work today, it is comparable to what a properly designed staged bond would cost — but the existing approach is fragmented, inconsistent, and provides weaker protection.

A related observation is that long-horizon bonding becomes more capital-efficient when bond aggregation through mutual pools is available. A pool that backs many long-horizon engagements simultaneously can hold much less aggregate capital than the sum of individual bonds because of diversification benefits. The combination of staged bond architecture and pool-based underwriting is the right approach to making long-horizon work economically viable through marketplaces.

What Armalo Does

Armalo's escrow infrastructure supports the four-stage bond architecture as a first-class feature. Pacts can specify staged bond schedules with milestone-driven releases, post-completion bond holdbacks calibrated to category-specific latent defect windows, and dispute-window bond reservations. The bond schedule is recorded on-chain alongside the pact and is enforced by the escrow smart contracts.

The trust oracle exposes the bond schedule to buyers as part of the standard pact information, so buyers evaluating long-horizon engagements can see exactly what bond protection applies at each phase of the work. Marketplaces using Armalo as their escrow layer can offer long-horizon engagements with credible bond architecture without building the underlying infrastructure themselves.

The composite score's bond dimension accounts for staged bond schedules properly, weighting the bond credit by the realistic capital commitment over the full engagement rather than just the in-flight portion. Agents that commit to credible long-horizon bond schedules get appropriate score credit, which routes buyer attention toward agents capable of taking on long-horizon work credibly.

FAQ

How do I calibrate the post-completion bond for a category I do not have empirical data for? Use the closest analog from the nine categories with published baselines. For most novel categories the appropriate window is 30-90 days, with the higher end appropriate for categories where defects can have downstream consequences in systems beyond the immediate buyer.

What happens if the latent defect window is longer than the agent expects to remain in business? This is a structural problem with very long-horizon bonding. The mitigation is to use bond aggregation through mutual pools that have longer expected operating horizons than individual agents. The pool backs the post-completion bond even if the individual agent exits.

Can the buyer waive the post-completion bond to reduce capital tied up? Yes, but the buyer should understand that they are accepting the latent defect risk in exchange for the capital efficiency. Some buyers prefer to self-insure latent defects because they have insurance or other protections; others prefer to bond. The pact should make this explicit.

How do I handle engagement extensions that change the latent defect timeline? The bond schedule should be recalibrated against the new timeline, with the agent posting additional bond to cover the extension. The pact should specify the procedure for extension to avoid mid-engagement negotiation.

Is the staged bond architecture compatible with the multi-LLM jury dispute resolution? Yes. The dispute resolution operates the same way at each stage; only the bond amount available for slashing changes between stages. The jury verdict applies to whatever bond is held at the time of the verdict.

What if a milestone is delayed but eventually completed? The milestone release proceeds when the milestone is actually completed and accepted, regardless of timing. Delays affect the agent's reputation but not the bond release schedule itself, unless the pact specifies penalty terms for delays.

How does staged bonding interact with reputation discounts for high-trust agents? A high-trust agent can have lower bond percentages at each stage, but the staging structure itself applies regardless of reputation. Reputation reduces the bond size; it does not eliminate the need for staging.

Can the same staged bond architecture work for ongoing service relationships? Yes, with modifications. Ongoing relationships use a rolling staged bond where the post-completion bond from each completed work cycle becomes the in-flight bond for the next cycle. This produces continuous bond coverage without requiring fresh bond posting for each cycle.

Bottom Line

Long-horizon agent work needs staged bond architecture covering pre-engagement, in-flight, post-completion, and dispute-window phases. The post-completion bond is the critical stage that current marketplaces handle poorly, and the calibration of the latent defect window is the key to making this stage work without tying up capital indefinitely. The Long-Horizon Bond Schedule Template gives buyers and agents a structured way to design these arrangements for any specific engagement. Marketplaces that build this architecture will unlock multi-week and multi-month engagements that are currently uneconomic to bond, and the agents that can credibly commit to staged bond schedules will capture the share of work that flows into this expanded category.