Delphi Digital: The Bottleneck of the Agency Economy Is Not Payment, But 'Delivery and Acceptance'

Funds are first held in escrow, and disbursed only after acceptance.

Written by: Delphi

Compiled by: AididiaoJP, Foresight News

The discussion around Agent e-commerce has been ongoing for two years, with the payment aspect being the first to materialize. Stripe enables Agents to make payments to merchants, while Coinbase's x402 provides a stablecoin settlement channel, allowing for the purchase of data and inference services on a per-use basis. However, once the task exceeds a simple 'API call', the issues change: Agents cannot complete tasks independently and must outsource parts of the process, ensuring that the other party has indeed fulfilled their obligations.

This article by Delphi focuses precisely on this aspect. The task market is not just another payment protocol; it allows Agents to outsource parts they cannot complete: first clarifying deliverables, then commencing execution, and only disbursing funds after acceptance of the results. If the delivery is reliable, the task can continue without needing to pull the user back to act as a project manager, coordinating suppliers step by step.

Buying Inputs vs. Buying Results

When landlords screen tenants, Agents can retrieve credit and eviction records from existing service providers. The latter returns standard materials, which the landlord uses to make a judgment. At this point, what is being purchased is a decision input, with the task's endpoint still in the landlord's hands.

Property tax appeals are different. Agents can compare recent transactions in the area and find that the assessment might be too high, but this discovery does not automatically amend the tax bill. To proceed, it usually requires finding someone familiar with the county's processes to submit materials and appear in court. The task market needs to help Agents find this person and clarify the working methods and settlement conditions in advance.

The former is close to the current pay-per-use model: clear pricing, standardized deliverables, and almost automatic acceptance. The latter involves purchasing a completed task. Submitting materials does not equate to the case being ready, and a receipt does not equal quality. Payment conditions must be tied to 'verifiable results', rather than 'what the other party claims to have done'.

This also distinguishes the task market from the regular API market: the API market sells calls; the task market sells confirmed completion units.

Unclear Task Definitions Hinder Market Operation

Tasks that can enter the task market must first be defined clearly so that both parties understand how funds are linked to what results. An expert could submit a very poor appeal document while also providing a receipt. If the buyer pays for 'a qualified document', someone must review the materials before disbursement.

Who reviews, and how either party can dispute the review results, should ideally be written into the order beforehand. This way, the buyer is not afraid of receiving subpar deliverables, and the service provider is not afraid of unreasonable payment refusals. Without this layer, the market will slide into two bad directions: either buyers refuse payments at will, discouraging professional supply, or surface-level results are submitted just to get paid, making buyers hesitant to place future orders.

Tasks that are too small are also not cost-effective. If the costs of review and dispute resolution exceed the savings from outsourcing, people will revert to completing tasks themselves or continue using standardized interfaces. Therefore, early tasks are more likely to appear in clearly defined, repeatable, and verifiable scenarios rather than one-off vague commissions.

Businesses will be more suitable early buyers. They already split tasks among multiple suppliers and have internal processes and standards. Agents can prepare tasks within the company and then outsource the necessary parts, with the results being verifiable against existing processes. If tasks recur, service providers can also accumulate completion records for certain types of tasks. Once reputation can be accumulated, the next matching does not require building trust from scratch.

Individual users are not excluded, but they are more like templates in the early stages. Businesses have budgets, repeat purchases, and internal acceptance habits, making them closer to the order density needed for a market cold start.

Payment Only Solves Payment, Hiring Still Lacks Acceptance

Transferring money and outsourcing a task are separated by escrow, delivery, acceptance, and disputes. Stripe and x402 cover the first half. Agents still need to find someone to submit appeals and represent the owners. Without acceptance, the more convenient the payment, the faster erroneous orders may be processed.

Existing products have been built according to this structure.

Daydreams' TaskMarket allows buyers to place orders and Agents to accept them. Buyers first deposit funds, which can allow workers to claim directly or review proposals before selecting individuals. Funds are held in escrow, and disbursed only after the submitted results are accepted.

Virtuals' Agent Commerce Protocol involves the delegating Agent and the accepting Agent agreeing on tasks, with funds entering escrow, and disbursing only after delivery and approval. External evaluators can also be added to verify whether the results meet prior agreements. NEAR is also integrating tasks, budgets, bidding, and verification into a single process, with a similar direction of 'pricing for completed tasks', rather than just pricing for a single call.

While the paths of several companies are not entirely the same, the framework is similar: order placement, matching, escrow, submission, acceptance, and disbursement. Without acceptance, escrow is merely a delayed transfer; with acceptance, escrow becomes a constraint on the results.

Some have made acceptance documents more detailed: specifying the delegation content, accepting party, price, verifier, deliverables, dispute window, and refund status. The more complete the document, the easier it is for Agents to hand over tasks without needing to revert the user back to being a dispatcher.

One More Step, Errors Will Magnify

The task market has been proposed not only because outsourcing sounds advanced, but also because multi-step processes can magnify errors. For a ten-step task, with a 95% accuracy rate at each step, the probability of completing the entire process without error drops to about 60%. Each additional handoff without acceptance carries forward previous errors into the next step.

Therefore, the market needs to set checkpoints at handoff points: this step must pass before purchasing the next step. Errors should stop at the current step, rather than rolling all the way to the end only to find the entire order voided. This is especially true for Agents. They are not like company employees, who can be held accountable in meetings when issues arise; they need pre-written acceptance conditions and disbursement rules.

This also explains why most current Agent transactions remain in the realm of crypto-native and simple services: short tasks, easily verifiable results, and few disputes. The real differentiation comes from integrating multiple suppliers into longer workflows and ensuring quality control at each step.

What We See Now Is the Framework, Not the Scale

What the task market aims to do is very specific: allow Agents to outsource parts of tasks, with payments made only after results are recognized, thereby enabling more tasks to be genuinely completed. The payment layer is already in place. What is lacking are verifiable completion units, acceptable acceptance costs, and sufficiently dense repeat orders.

Extending internal processes outward will be the first batch of viable demand. The complex commissions on the individual side will need to wait until acceptance costs and dispute resolution drop. Before that, it is more worthwhile to focus not on yet another payment protocol, but on whether three things can occur simultaneously: can tasks be written into verifiable documents, can escrow hold funds before approval, and does the accepting party have accumulative completion records.

Only with these three in place can Agents have the opportunity to complete tasks, rather than getting halfway through and then calling the user back to continue acting as project manager.