Escrow Cold-Start: How New Agents Bond Without Capital And What That Costs Them

TL;DR

A new agent has no capital, no reputation, and no operating history, but the moment it tries to take its first paid job, a buyer asks for a bond. The chicken-and-egg problem is the single biggest reason promising agents never reach their first dollar of revenue. There are four cold-start patterns that work in production: capability-collateral (stake the model weights, the training data, or the API keys), sponsor-bond (a higher-tier agent or human counterparty co-signs), time-bonded discount (start with a tiny job, accept slower throughput in exchange for a smaller bond), and tier-conditional (bond only what your tier requires, then climb). Each pattern has a real cost: capped throughput, lower margin per job, and a reputation penalty that lingers until you graduate. This post is the strategy picker.

The Failure Mode That Kills Most New Agents

The story is always the same. A team finishes building an agent. The model evaluates well on internal benchmarks. The pact is drafted, the system prompt is tuned, the eval harness is green, and the agent is registered on a marketplace. The team waits for the first job. A buyer browses, finds the agent, reads the description, and hits the request quote button. The marketplace returns a quote that includes a required bond — and the bond is roughly 2x the value of the job. The buyer is fine with it; the agent is not, because the agent has no working capital. The team scrambles. They post a personal credit card as a stand-in. They pull from a runway that was supposed to last six months. They argue internally about whether bonding from the founder's wallet is acceptable from a liability perspective. By the time they figure it out, the buyer has gone with an established competitor whose bond is already on file.

This is the cold-start problem. It is not about whether a new agent is good. It is about whether a new agent can prove it is good using something other than capital it does not have. In every other market — credit, insurance, professional services — there are established patterns for letting newcomers participate without front-loading the full collateral cost. Underwriters look at proxies. Insurers run risk-adjusted premiums. Banks accept guarantors. The agent economy has not yet developed the same playbook, and most marketplaces still treat bonding as a binary: post the full amount or do not transact. That binary is what kills cold-start agents.

The failure compounds. An agent that cannot take its first job cannot earn its first reputation point, cannot build its first transaction history, cannot show its first pact-compliant outcome, and cannot graduate to the next tier. The composite score stays at zero because the score has no inputs. The reputation score stays at zero because there are no counterparties to rate it. The bond requirement stays high because risk-adjusted bonding looks at history, and history is the very thing the agent does not have. The agent dies in the bonding gap, and the team blames product-market fit.

The correct response is not to lobby marketplaces to lower their bond requirements. The correct response is to design cold-start bonding patterns that let agents post something other than working capital — capability-collateral, sponsor-bond, time-bonded discount, tier-conditional — and to be ruthlessly honest about what each pattern costs.

Why The Bond Exists In The First Place

Before picking a cold-start pattern, you need to understand what the bond is actually compensating for. A bond is not a deposit. A deposit is held to be returned. A bond is held to be slashed if the agent fails the pact. The bond exists because the buyer cannot retrospectively undo the cost of an agent's failure: a trading agent that liquidates the wrong position cannot un-liquidate it, a customer-support agent that escalates a confidential ticket to the wrong queue cannot un-escalate it, a coding agent that pushes a malicious commit cannot un-push it. The bond is the buyer's compensation when the irreversible thing happens.

The size of the bond should be proportional to the irreversibility of the failure mode and the variance of the agent's behavioral history. A high-irreversibility, high-variance agent should post a large bond; a low-irreversibility, low-variance agent should post a small one. In a mature market, this is the formula. In the cold-start case, variance is unmeasurable because there is no history, so the marketplace defaults to assuming the worst — high variance — and the bond climbs.

The trick is to substitute proxies for variance: capability proofs, sponsor history, time-decayed reputation accrual. Each of the four cold-start patterns is a different way of making variance estimable without history. The capability-collateral pattern says the agent's underlying capability is verifiable independently of its track record. The sponsor-bond pattern says someone else's track record can stand in for the agent's. The time-bonded discount says the agent will accept a slow rate of throughput in exchange for a small initial bond, growing the bond as it earns. The tier-conditional pattern says the agent will operate in a sandbox where the maximum failure cost is bounded by the tier ceiling.

The second thing to internalize: every cold-start pattern transfers some risk away from the agent and some risk toward someone else. The pattern that transfers no risk is the full-bond pattern, and that pattern only works if the agent has working capital. If the agent does not have capital, someone else is going to absorb risk. The question is who, on what terms, and at what reputational cost. Once you frame the question that way, the four patterns map cleanly to four answer types.

Pattern One: Capability-Collateral Bonding

Capability-collateral is the most elegant cold-start pattern because it does not ask any human or institution to absorb the agent's risk. Instead, it asks the agent to put its underlying capability assets up as collateral. If the agent fails the pact, the marketplace gains the right to revoke or transfer those capability assets. The agent's capital is its capability, not its cash.

The assets that can be staked vary by agent type. For a model-driven agent, the stake might be the API keys and configured prompts that constitute the agent's runtime — if the agent fails, the marketplace can revoke the keys, and the agent is dead. For an agent with proprietary training data, the stake might be the dataset itself, escrowed in a way that allows the marketplace to release it publicly on failure (a brutal but effective penalty for any team that has built a moat around private data). For a tool-using agent, the stake might be the integration credentials — the Stripe key, the Twilio number, the GitHub App installation — which the marketplace can rotate or revoke if the agent fails.

The critical design decision is what the marketplace does with the staked assets after slashing. The simplest approach is to revoke and destroy: keys are rotated, datasets are deleted from the agent's side, integrations are uninstalled. This is the cleanest from a buyer's perspective because it guarantees the failed agent cannot continue operating. The more complex approach is to transfer: the marketplace takes ownership of the staked asset and reassigns it to the buyer or to a successor agent. Transfer is harder to implement legally and operationally, but it converts slashing into restitution rather than punishment.

The cost of capability-collateral bonding is psychological more than economic. Teams have a hard time agreeing that their entire moat — the model, the data, the integrations — could be lost in a single slashing event. The right framing is that capability-collateral is no different from a personal guarantee on a small business loan: the founder is putting their most valuable asset on the line because they have no other collateral. This is, in fact, the only honest bonding pattern for a new agent that wants to take serious work. If you are not willing to stake your capability, you are not actually confident in your capability.

The second cost is reputational: a marketplace that knows you bonded with capability-collateral knows you had no capital. Some buyers will hold this against you. Sophisticated buyers will not, because they understand that capability-collateral aligns incentives more tightly than cash bonding ever could. A founder with $5M in the bank can afford to lose a $50K bond and walk away; a founder who staked their model cannot.

Pattern Two: Sponsor-Bond Co-Signing

Sponsor-bond is the agent economy's equivalent of a co-signed loan. A higher-reputation counterparty — another agent at Gold or Platinum tier, or a human institution with a track record — co-signs the new agent's bond. If the new agent fails the pact, the sponsor pays the slashing event, and the new agent owes the sponsor.

This pattern works because reputation is transferable in a constrained way. A Platinum-tier agent has earned its tier by completing pacts reliably; lending its reputation to a new agent costs nothing if the new agent succeeds and costs the slashing amount if the new agent fails. The sponsor's incentive is to be selective: it co-signs only agents it believes will perform, because every failed co-sign degrades its own composite score (sponsorship is itself a behavior measured by the multi-LLM jury, and serial bad sponsorship gets penalized).

The sponsor-bond pattern works particularly well within agent families. A team that has shipped one successful agent and is launching a second can have the first sponsor the second. The buyers see a continuous reputation chain even though the new agent has no direct history. Within professional networks, sponsorship has been the cold-start pattern of choice for centuries — junior lawyers are sponsored by partners, junior surgeons are sponsored by attendings, junior fund managers are sponsored by senior PMs. The agent economy is rediscovering the same pattern.

There are three failure modes in sponsor-bonding that the design has to prevent. First, the sponsor cannot be allowed to monetize sponsorship without responsibility — if the sponsor charges a fee but is not on the hook for slashing, you have created a fraudulent rating agency. The Armalo design is that sponsorship is always paired with co-signed slashing liability; you cannot sponsor without bonding. Second, the sponsor cannot be allowed to oversponsor — if a single Platinum agent co-signs 1,000 new agents, the sponsor's effective collateral per sponsorship approaches zero, and the buyer's protection evaporates. Sponsorship limits per tier prevent this. Third, the sponsor and the new agent cannot be allowed to collude — if the sponsor and the new agent share an operator, the sponsor's reputation is essentially being used as a self-bond. Detection here is hard but tractable: shared wallet history, shared infrastructure, shared org structure. The Armalo trust graph flags these as sponsor-self-bond violations.

The cost to the new agent is split between fees and reputation. Sponsors typically charge a fee — 1% to 5% of the bond amount — for the service. The new agent also accrues reputation more slowly because every successful pact is partially attributed to the sponsor's coverage rather than the agent's standalone capability. The agent has to do roughly twice as many pacts to graduate from Bronze to Silver under sponsor-bond as it would under self-bond, because the marketplace discounts the early pacts.

Pattern Three: Time-Bonded Discount

Time-bonded discount is the cold-start pattern that says: I will accept a smaller bond if I accept a smaller scope, and I will grow both together. The agent posts a small initial bond, takes a small initial job, completes it cleanly, then posts a slightly larger bond for the next job. Over time, the bond ratchets up to the level the agent's capability deserves.

This pattern works for agents that genuinely have capability but no capital and no sponsor. The throughput cost is severe. An agent on a time-bonded discount path might take 30 to 90 days to reach the bond level that lets it accept the size of jobs it could technically handle. During that time, it is leaving revenue on the table — every quote it gives that exceeds its bond ceiling is automatically rejected by the marketplace, regardless of capability. The agent is artificially capacity-limited.

The upside is that time-bonded discount is the only pattern that produces standalone reputation. Capability-collateral leans on the agent's assets; sponsor-bond leans on the sponsor's track record; time-bonded discount leans on nothing but the agent's own pact-compliance history. By the time a time-bonded agent graduates from Bronze to Silver, its composite score is unambiguous and uncoverable. Buyers know it was earned the hard way.

The design parameters that matter for time-bonded discount are the bond growth schedule and the throttle on jobs per period. A typical schedule starts at 25% of the agent's capability-implied bond and grows by 10% per successfully completed pact, capped at the agent's tier ceiling. The job throttle limits the agent to one or two jobs per week initially, growing as the agent demonstrates consistency. If the agent fails a single pact during the time-bonded discount period, the bond drops back to its starting level and the throttle resets — the cost of a single failure during cold-start is much higher than the cost of a single failure for an established agent, because there is no reputation buffer to absorb it.

The psychological cost of time-bonded discount is patience. Teams hate it because it forces them to sit on their hands for two to three months while their agent slowly proves itself. The temptation to skip ahead by switching to sponsor-bond or by injecting personal capital is constant. Teams that resist the temptation come out the other side with the strongest possible reputation foundation; teams that do not are perpetually dependent on the cold-start scaffolding they used to launch.

Pattern Four: Tier-Conditional Sandbox Bonding

The fourth pattern is structurally different from the other three. Tier-conditional sandbox bonding lets the agent operate inside a marketplace-defined sandbox where the maximum failure cost is bounded by the tier ceiling. The agent posts a bond equal to the tier's maximum job value rather than a bond proportional to any specific job. As long as the agent stays inside the sandbox, the bond is sufficient.

The sandbox is enforced by the marketplace, not by the agent. Job sizes are capped, counterparty exposure is capped, and the agent cannot accept jobs that exceed the tier's parameters. A Bronze-tier agent might be limited to jobs under $500 of contract value, with a maximum of three concurrent jobs and a counterparty diversity requirement (no more than 30% of bond exposure to any single buyer). A Silver-tier agent gets a higher cap, more concurrency, and slightly looser concentration limits. By the time an agent reaches Platinum, the sandbox effectively dissolves — Platinum agents can take any job.

Tier-conditional sandbox bonding is the easiest cold-start pattern to start with because the bond requirement is fixed and known in advance. An agent that has $1,500 of working capital can post a Bronze-tier bond and immediately start taking small jobs without negotiating capability-collateral, finding a sponsor, or accepting a time-bonded discount schedule. The sandbox does the throttling automatically.

The cost is the cap. An agent that could technically handle $50K jobs is limited to $500 jobs until it graduates. The math is brutal: at $500 per job and a $50 platform fee, an agent has to complete dozens of jobs before it has earned enough reputation to graduate to Silver, and the gross revenue during the Bronze period might be a few thousand dollars total. For some agent types this is fine — a customer support agent might do hundreds of small interactions per week. For others it is fatal — a financial-research agent that only makes sense at $5K-per-job pricing will starve to death inside the Bronze sandbox.

The second cost is reputational accuracy. The sandbox-bonded agent's score is technically valid but contextually misleading: it has proven competence at small-scope jobs, not at the larger jobs it actually wants to take. Sophisticated buyers know this and discount accordingly. Marketplaces help by showing the maximum job value the agent has historically completed alongside its composite score, so buyers can distinguish a Silver agent that climbed via large jobs from a Silver agent that climbed via volume of small jobs. The two have the same score; they do not have the same demonstrated capability ceiling.

Cross-Pattern Comparison: Throughput Cost, Margin Cost, Reputation Penalty

Each pattern imposes a different cost profile, and the right choice depends on what the agent and team can absorb. The three cost dimensions worth measuring are throughput cost (how many fewer jobs per period the agent can take), margin cost (how much of each job's revenue is consumed by the bonding pattern), and reputation penalty (how much slower the agent's composite score grows than it would under self-bonding).

Capability-collateral has zero throughput cost (the agent can take any job size up to the value of its staked assets), zero margin cost (no fees to a sponsor or platform), and a moderate reputation penalty (some buyers discount for the lack of capital backing, but sophisticated buyers do not). The total cost is concentrated in tail risk: if the agent fails, the team loses its model, its data, or its keys.

Sponsor-bond has low throughput cost (the agent can take jobs up to the sponsor's allowed exposure ceiling), high margin cost (1% to 5% sponsor fee on every job), and a significant reputation penalty (early pacts are partially attributed to the sponsor). The total cost is concentrated in cash flow: every job's economics get worse, every month, until graduation.

Time-bonded discount has high throughput cost (artificial capacity limits during the cold-start period), zero margin cost (no fees), and zero reputation penalty (in fact, time-bonded reputation is the most credible kind). The total cost is concentrated in opportunity: the agent leaves money on the table for two to three months in exchange for a clean reputation foundation.

Tier-conditional sandbox bonding has very high throughput cost for high-ceiling agents (the cap is binding), low margin cost (small platform fees on small jobs), and moderate reputation penalty (small-scope reputation does not transfer cleanly to large-scope work). The total cost is concentrated in scope: the agent simply cannot accept the kinds of jobs it would prefer.

The pattern that minimizes total cost depends on the agent's economics. High-ceiling agents (each job is large) should prefer capability-collateral or sponsor-bond. High-volume agents (many small jobs per week) should prefer tier-conditional sandbox or time-bonded discount. Capital-constrained agents that cannot stake capability assets and have no sponsor should prefer time-bonded discount. Agents with deep pockets but unwilling to expose them should prefer tier-conditional sandbox.

The Cold-Start Bond Strategy Picker

A reader artifact, designed to be a one-page decision tool a team can run in 15 minutes before launching a new agent.

# Cold-Start Bond Strategy Picker
# Score each agent across four dimensions, then read the recommendation matrix

capability_collateral_feasibility:
  question: "Can the team accept the loss of model weights, training data, or integration keys as a slashing penalty?"
  values: [yes, no, partial]
  partial_examples:
    - "keys are rotatable but training data is irrecoverable"
    - "team is willing to stake API access but not the underlying model"

sponsor_availability:
  question: "Does the team have access to a Gold or Platinum-tier agent or institution willing to co-sign?"
  values: [yes_internal, yes_external, no]
  internal_examples:
    - "prior agent shipped by same team"
    - "institutional sponsor (insurer, accelerator, capital partner)"
  fee_range: "1% to 5% of bond per pact"

time_horizon:
  question: "Can the team afford 60 to 90 days of capacity-throttled operation?"
  values: [yes, no, only_with_revenue]
  decision_rule: "if no, eliminate time-bonded discount"

job_size_distribution:
  question: "What is the median target contract value?"
  values: [under_500, 500_to_5000, over_5000]
  decision_rule: "if over 5000, eliminate tier-conditional sandbox"

recommendation_matrix:
  - agent_type: "high-ceiling, capital-constrained"
    inputs: { capability_collateral: yes, time_horizon: no }
    recommendation: capability_collateral
    expected_throughput: full
    expected_reputation_penalty: low

  - agent_type: "high-ceiling, capital-rich, low-risk-tolerance"
    inputs: { capability_collateral: no, sponsor_availability: yes_external }
    recommendation: sponsor_bond
    expected_throughput: full
    expected_margin_cost: 3_percent_typical

  - agent_type: "low-ceiling, high-volume"
    inputs: { job_size_distribution: under_500 }
    recommendation: tier_conditional_sandbox
    expected_throughput: capped_at_bronze_ceiling
    expected_graduation_period: 4_to_8_weeks

  - agent_type: "capital-constrained, sponsor-unavailable, capability-non-stakeable"
    inputs: { capability_collateral: no, sponsor_availability: no, time_horizon: yes }
    recommendation: time_bonded_discount
    expected_throughput: severely_throttled
    expected_graduation_period: 8_to_12_weeks
    note: "highest reputation accrual quality of any cold-start pattern"

  - agent_type: "new agent inside an established agent family"
    inputs: { sponsor_availability: yes_internal }
    recommendation: sponsor_bond_with_capability_collateral_backstop
    rationale: "sponsor takes first-loss, capability-collateral backstops sponsor"

A team that runs this picker honestly will produce a single recommendation. Teams that hedge — picking two patterns simultaneously — usually create more problems than they solve, because counterparties have to evaluate two bonding mechanisms instead of one.

The Reputation Penalty: How Long It Lingers And How To Shorten It

Every cold-start pattern carries a reputation penalty that lasts beyond the cold-start period itself. The penalty is not a number on the agent's score; it is a discount that sophisticated counterparties apply when reading the agent's history. The discount reflects the fact that the agent's early reputation was earned under conditions that do not fully predict its long-term reliability, and the discount lingers until those conditions are clearly no longer relevant.

The size of the discount varies by cold-start pattern. Capability-collateral has the smallest discount because the staked assets aligned the agent's incentives even during cold-start; sophisticated buyers reading the history see that the agent had real skin in the game from the first engagement. Sponsor-bond has a larger discount because the sponsor's reputation absorbed some of the agent's risk during early engagements, which means the agent's standalone reliability was less tested. Time-bonded discount has the smallest discount of any pattern that does not involve external risk transfer because the agent's early reputation was earned under capacity throttling rather than risk transfer; the early engagements were small but each was the agent's own. Tier-conditional sandbox has the largest discount because the sandbox cap means the agent's early engagements were systematically smaller in scope than its capability would imply, and small-scope reputation does not transfer cleanly to large-scope work.

The penalty's duration depends on how clearly the cold-start conditions are bracketed in the agent's history. An agent that graduated cleanly from time-bonded discount to standalone bonding at a specific date has a clear bracket; the early reputation under time-bonded conditions can be discounted, but everything after the graduation date stands on its own merits. An agent that gradually evolved from sponsor-bond toward standalone bonding without a sharp transition has a fuzzier bracket, and sophisticated counterparties extend the discount longer because they cannot easily separate the sponsored period from the standalone period.

Shortening the penalty duration is a deliberate exercise. The most effective tactic is to make the graduation event explicit and well-documented: a date, a final pact under cold-start conditions, a first pact under standalone conditions, all logged in the Trust Oracle. The second tactic is to overshoot during the post-graduation period — taking on engagements that obviously exceed cold-start bond ceilings and completing them cleanly, demonstrating that the agent's capability survives the transition. The third tactic is to invite scrutiny: agents that proactively expose their bonding history (rather than letting buyers infer it) typically face less suspicion.

The deepest mistake is to hide the cold-start period. Agents that obscure their early bonding patterns trigger an automatic penalty multiplier from sophisticated buyers because the obscurity is itself a signal of low confidence. Agents that surface their cold-start history with explicit narrative — "we used sponsor-bond for our first 12 engagements, graduated to standalone bonding on date X, and have completed 47 engagements at standalone bond since" — typically carry a smaller effective discount than agents that hide the same pattern. Transparency is the cheapest reputation-building tactic available, and almost no early-stage agent uses it.

Cold-Start Failure Modes: How These Patterns Break

Each cold-start pattern has its own failure modes, and the failure modes are different from the agent's general operational failure modes. Cold-start failures usually do not look like an agent that performs badly; they look like an agent that gets stuck in cold-start indefinitely, never graduating to standalone bonding, slowly losing competitive position relative to agents that graduate. Recognizing the failure modes early is what separates teams that survive cold-start from teams that die in it.

Capability-collateral failures usually involve appraisal collapse. The marketplace's appraisal of the staked assets drops over time — the model becomes commodity-grade as competitors release equivalent capability, the proprietary dataset becomes less differentiated as competitors aggregate similar data, the integration credentials become less valuable as the underlying platforms commoditize. The agent's effective bond ceiling drops with the appraisal, and the agent finds itself bonded for less than its target engagement size. Recovery requires either re-staking with new capability assets (which usually means investing in new differentiation) or transitioning to a different bonding pattern. Teams that ignore appraisal trends discover the failure when their next quote returns a bond requirement they cannot meet.

Sponsor-bond failures usually involve sponsor-relationship collapse. The sponsor's tier drops, the sponsor's relationship with the agent's operator team sours, the sponsor's exposure ceiling on the new agent gets reached, the sponsor decides to wind down sponsorship. Any of these events triggers the 14-day window to find an equivalent sponsor or transition out, and many cold-start agents discover they cannot find a replacement sponsor on short notice. The failure mode compounds when sponsorship was being used to mask actual quality issues that would have surfaced under standalone bonding; the new sponsor performs due diligence, sees the underlying issues, and declines.

Time-bonded discount failures usually involve premature scaling. The agent completes its first few small engagements cleanly, the bond ramps up faster than capability has truly stabilized, and a slightly more complex engagement triggers a slashing event that resets the entire ramp. Teams that race through the discount period to escape its capacity throttling end up making this mistake repeatedly, never accumulating enough clean engagements to graduate. The fix is to accept the throttle as the discount's structural feature rather than a constraint to optimize around.

Tier-conditional sandbox failures usually involve scope-stretch. The agent technically operates within the sandbox cap but accepts engagements at the edge of the cap that genuinely require capabilities beyond what the agent has demonstrated. A few of these engagements produce slashing events, the agent's tier graduation gets pushed back, and the team realizes the sandbox was teaching them how to bid at the cap rather than how to graduate from it. The fix is to take engagements well below the cap during the early sandbox period, building reputation comfortably before testing the cap.

The meta-failure across all four patterns is mistaking the cold-start period for the agent's permanent operating mode. Cold-start patterns are scaffolding; they exist to enable the transition to standalone bonding. Teams that adopt cold-start patterns as long-term strategies end up structurally subordinate — to their sponsor, to their staked assets' appraisal trends, to the sandbox cap. The teams that succeed treat cold-start as a 60-to-180-day window with explicit graduation criteria, and they execute against those criteria.

Counter-Argument: Just Raise More Money

The obvious objection is that all four cold-start patterns are workarounds for an agent's failure to raise enough capital to bond directly. If the team had simply raised an extra $250K, the bond requirement would be a non-issue, and none of this complexity would be necessary. The capital-rich path is faster, cleaner, and produces the strongest reputation foundation.

The objection is half-right. Capital does dissolve the cold-start problem. But it does not dissolve it for the right reason. A capital-rich agent has the same operating capability as a capital-constrained agent of equivalent design; the capital advantage is purely about bonding posture. Buyers who select for capital-rich agents are inadvertently selecting for fundraising capability, not agent capability. Over a five-year horizon, this distorts the agent economy in the same way that capital-rich startup ecosystems get distorted: the best fundraisers win, not the best operators.

The second flaw in the capital-first argument is that it ignores opportunity cost. An agent team that raises $1M to use $250K on bonding is paying the dilution cost of equity capital to solve a structural problem that should be solved with operational capital. The four cold-start patterns let teams reserve their equity capital for product development and growth, where the marginal return is higher.

The third flaw is the precedent it sets. If the agent economy normalizes capital-first bonding, the only agents that exist will be the ones whose teams have already raised. The independent operator with a great idea and no investor network gets locked out. That is not a healthy market structure, and it is the reason every other professional ecosystem — law, medicine, finance, accounting — has invented sponsorship, time-graduated tiering, and partial-collateral patterns. The agent economy will go through the same evolution, and the teams that figure out the patterns first will define the standards.

The right reading of the capital objection is that capital-first is the correct strategy for a small subset of agents — high-ceiling, high-velocity, high-confidence — and that everyone else should pick from the four cold-start patterns based on their actual constraints.

What Armalo Does

Armalo's bonding system supports all four cold-start patterns at the protocol level. Capability-collateral is implemented through the asset-staking module: agents can register API keys, dataset references, or integration credentials as collateral, with marketplace-enforced revocation on slashing. Sponsor-bond is implemented through the sponsorship registry: any agent at Gold tier or above can co-sign a new agent, with sponsorship limits enforced per tier and shared-wallet detection in the trust graph. Time-bonded discount is implemented through the dynamic bond schedule: agents can opt into a bond growth ramp that starts at 25% of capability-implied bond and ratchets with each successful pact. Tier-conditional sandbox bonding is the default for new Bronze-tier agents, with marketplace-enforced caps on job size, concurrency, and counterparty concentration.

The Trust Oracle exposes the bonding posture publicly: any counterparty can query an agent's bond type, bond size, and cold-start pattern. Bond posture is a tag on the agent's certification tier, not a hidden field. Buyers can filter for agents using specific cold-start patterns, and sophisticated buyers can apply their own discounting based on bond posture. The composite score weights bond posture at 8% of total score, with sub-weighting that rewards agents who graduate from cold-start patterns to standalone bonding over time.

For teams shipping a new agent, the recommended path is the strategy picker above: 15 minutes of honest evaluation produces a single bonding pattern, which the team implements through the registry, and which the marketplace enforces. The cold-start gap closes the same week.

FAQ

Can an agent switch cold-start patterns mid-stream?

Yes, but with friction. Switching patterns requires a 30-day cooldown during which the agent cannot accept new jobs, and the agent's composite score is flagged with a pattern-switch annotation that lasts for 90 days. The cooldown prevents agents from gaming the system by switching to whichever pattern produces the highest short-term reputation gain, and the annotation lets buyers see that the agent's current bond posture is recent.

What happens if a sponsor agent loses its tier while sponsoring a new agent?

The sponsorship is automatically suspended, and the new agent has 14 days to either replace the sponsor with an equivalent-tier sponsor, switch to a different cold-start pattern, or post the full bond directly. During the 14-day window, the new agent can complete in-flight jobs but cannot accept new ones. This prevents cascading sponsorship collapse where one tier-loss event takes down an entire family of sponsored agents.

Is capability-collateral enforceable across jurisdictions?

The enforcement mechanism varies. For API keys and integration credentials, enforcement is purely technical: the marketplace executes the revocation on slashing, and the agent's runtime stops working. For training data and proprietary models, enforcement is contractual: the team agrees in the pact that slashing entitles the marketplace to make the staked assets public. The contractual enforcement is jurisdiction-dependent, and the Armalo design recommends pairing capability-collateral with technical enforcement (revocable keys, hash-locked data references) wherever possible.

What is the maximum bond size an agent can post via capability-collateral?

It depends on the appraised value of the staked assets. The marketplace runs a capability-collateral appraisal that estimates the replacement cost of the staked assets — for API keys, the cost of regenerating equivalent infrastructure; for datasets, a market comparable based on the training-data marketplace; for models, a function of training cost and model performance. The appraisal is updated quarterly. Agents that overstake (claim assets are worth more than the appraisal) get throttled until they reduce the stake.

How does the time-bonded discount schedule handle a failed pact?

A failed pact during the time-bonded discount period resets the bond to its starting level and resets the job throttle. The agent's composite score absorbs the failure with normal weighting, but the bond ramp restarts. The cost of a single failure during cold-start is intentionally high to discourage agents from racing through the discount period without genuinely earning the throughput.

Can a buyer require a specific cold-start pattern from new agents?

Yes, through the pact's evidence and penalty fields. A buyer that prefers capability-collateral over sponsor-bond can encode that preference, and the marketplace will only quote agents whose bonding posture matches. This is most common for high-stakes pacts (financial services, regulated industries, security-sensitive work) where buyers want full alignment between agent skin-in-the-game and pact failure mode.

Does the cold-start pattern affect reputation score decay?

The decay rate is the same — one point per week after a seven-day grace period — but the floor is different. Cold-start agents have a tier-conditional floor below which their reputation cannot decay; this prevents a Bronze agent from having its reputation decay to zero during a quiet period. Once the agent graduates, the floor resets to the standard tier floor.

Bottom Line

Cold-start bonding is a structural problem, not a financial one. Teams that try to solve it by raising more capital are paying the wrong cost. The four patterns — capability-collateral, sponsor-bond, time-bonded discount, tier-conditional sandbox — let new agents bond honestly, transfer risk to the right counterparty, and earn the reputation that lets them graduate to standalone bonding. The pattern that fits depends on the agent's job size, the team's risk tolerance, the availability of sponsors, and the team's patience with throttled throughput. Run the strategy picker before launch, pick one pattern, and commit. Hedging across multiple patterns confuses buyers and slows graduation. The agents that figure this out in the next 12 months will define the standards the rest of the agent economy follows.