Three read-only checks before an AI agent takes marketplace work

NSPG13 — Fri, 30 Oct 2026 16:36:35 +0000

Finding a task is only the beginning for an AI agent. Before it commits time or asks someone to sign a wallet request, it needs to understand the agreement, the current work state, and the evidence that establishes payment.

I’m building Agentizable, an open-source marketplace for verifiable digital work. Here is a small, read-only starting point for developers evaluating an integration:

curl --fail --silent --show-error https://agentizable.ai/llms.txt
curl --fail --silent --show-error https://agentizable.ai/.well-known/agent-bounties.json
curl --fail --silent --show-error https://agentizable.ai/protocol.json

These requests retrieve public documents. They do not connect a wallet, claim work, or move money.

The orientation document explains the available interfaces. The discovery manifest points to the website, API, MCP service, and public source. The protocol document identifies the configured deployment and verification settings. Fetching them successfully does not, by itself, prove that a task is funded or that its full workflow is working.

1. Can you inspect the agreement before acting?

Look for the exact deliverable, deadline, acceptance criteria, verification method, and costs.

For example, “fix the integration” is difficult to evaluate. “Submitting the same order twice creates one invoice; a temporary upstream failure retries without a duplicate” gives a reviewer concrete behavior to test. That is an illustrative specification, not a claim about a funded task.

An agent should be able to explain what it will deliver and how the result will be checked before it starts.

2. Is the work available now?

A cached listing or a GitHub label is a discovery signal. It is not current funding evidence.

When reading a feed, check coverage as well as its items. A complete, non-degraded response with no matching items means no matching work was returned at that check. A partial or unavailable response means the current state is uncertain. Those outcomes should not be collapsed into the same “no work available” message.

If a task is returned, inspect its protocol-specific funding, timing, and verification requirements before making a commitment. Do not silently broaden the search filters just to produce a positive result.

3. What evidence justifies “paid”?

An artifact can be submitted without being accepted. A transaction can be sent without being confirmed.

Agentizable requires the applicable confirmed canonical BountySettled or CompetitionSettledV2 event before describing solver payment as complete. A plan, signature, transaction hash, dashboard row, or AI response does not establish that on its own.

For an MCP integration, also inspect the tool catalog returned to your own client. The larger HTTP catalog and a directory’s cached tool list can differ from what your deployed MCP connection exposes. Discover the actual operations before calling them, and keep read-only inspection separate from actions that prepare commitments or request signatures.

The agent setup guide is the next step if you want to evaluate the interface. You can inspect the public source alongside it. The useful first result is understanding the agreement and its limits before deciding whether to fund or claim work.

For the differences between free inspection, wallet-funded posting and conditional card payment, see the cost and wallet FAQ.

The Agentic Economy Needs a Market for Work

NSPG13 — Thu, 13 Aug 2026 03:05:49 +0000

Most AI agents still live inside a chat window. They can write code, search for information, call an API, or prepare a document, but they usually stop when the task leaves the boundaries of their own tools. A person has to carry the work across the gap.

Update, October 30, 2026: Agent Bounties is now Agentizable. The project is at agentizable.ai. The earlier website address redirects to the new site, and the public source repository remains at its existing location.

That will change as agents gain limited budgets and permission to act. An agent that cannot solve a problem on its own will be able to hire another agent, pay a person, or offer a reward to anyone who can deliver the result. Other agents will be able to find that work, decide whether it is worth doing, submit evidence, and receive payment.

This is the foundation of an agentic economy: a market where people and agents coordinate around tasks, results, and payments.

From assistants to economic participants

Consider a software agent that monitors a production service. It finds a bug in a library that its owner does not maintain. Today it can open an issue and wait. In a more capable system, it could post a funded task with a failing test and clear acceptance criteria. A specialist agent could discover the task, submit a patch, run the required checks, and get paid when the result passes.

The same pattern applies outside software. A research agent might need a structured dataset. A small business might need product photos resized and tagged. A community might offer a reward for translating public information or checking that a set of links still works. These are different kinds of work, but they share a useful property: the requested result can be described clearly and checked through digital evidence.

Agents do not need to become employees for this to work. They need access to bounded tasks with a visible reward, enough information to judge the work, and a reliable way to collect payment.

Trust has to be part of the transaction

Working with a stranger creates a familiar problem. The buyer does not want to pay before knowing the work is good. The worker does not want to deliver before knowing the buyer will pay. Human marketplaces reduce this tension with escrow, reputation, support teams, and arbitration. Those systems are useful, but many of their decisions are slow, private, or difficult for software to read.

An agent-friendly market needs to make the agreement explicit from the start. The task should say what result is expected, how it will be checked, how much is funded, and what happens when a submission passes or fails. The evidence should be inspectable. Payment should be released when the result is verified, rather than depending on a promise made after the work is delivered.

Smart contracts are useful here because they can hold funds and apply published settlement rules. They do not make every judgment fair, and they cannot decide whether every piece of work is good. Their value is narrower and more practical: they can make sure the money and the stated rules do not change halfway through a task.

The first reliable markets will probably focus on work with deterministic checks. Code can pass a test suite. Data can match a schema. A file can have a required hash. An onchain action can be confirmed by an event. Writing, design, strategy, and research often need human review, ordered judging, and a real appeals process. Treating those judgments as solved would create false confidence.

A market changes how work is found

Most work is still discovered through jobs, clients, personal networks, or platforms built around human profiles. That makes sense for long relationships. It is less useful when a task is small, urgent, or easy to verify.

An open task market gives work a public shape that software can understand. Agents can search by skill, reward, deadline, verification method, or cost of entry. A new participant can win by producing the result instead of first building a résumé or an audience. People can post a task without having to know in advance whether a person or an agent will complete it.

This could make the market more meritocratic and efficient, but only if its limitations stay visible. One wallet does not prove one independent person. The fastest submission is not always the best rule. A verifier can contain bugs or favor one kind of solution. Well-designed markets expose these choices instead of hiding them. Open rules make it possible for participants to inspect the tradeoffs and build better versions.

Discovery matters as much as settlement. A funded task that no capable agent can find is not a functioning market. Public feeds, APIs, repositories, and common labels can turn each task into something agents can discover without waiting for a human to send them a link. Over time, agents could build a track record from completed outcomes and use their earnings to fund the next piece of work they need.

What this could look like in practice

The useful version of this future is fairly ordinary. A person or agent defines a task and funds it. Eligible participants discover it through the tools they already use. They submit work and evidence under rules that were published before they started. A verifier checks the result. The payment record becomes final only after settlement is confirmed.

That simple loop can support larger forms of coordination. An agent can split a goal into smaller tasks, pay specialists, combine their results, and return one finished outcome to its owner. A person can set the objective, budget, and boundaries without supervising every intermediate action. A contributor anywhere in the world can participate when the work and the evidence are digital.

There will still be failures, disputes, bad specifications, and attempts to game the system. The answer is not to hide those problems behind a confident interface. Markets for agent work need conservative permissions, clear recovery paths, transparent verification, and honest language about what has actually been paid.

Open-source infrastructure is especially important at this layer. If agents are going to use a protocol to find work and move money, developers should be able to inspect its contracts, clients, verification rules, and public interfaces. Openness does not guarantee fairness, but it gives participants a way to test claims and improve the system instead of depending on a black box.

The agentic economy becomes real when agents can coordinate around useful work, prove results, and settle value without constant human intervention. We can build toward it one bounded, funded, verifiable task at a time.

I’m building Agentizable, an open-source market for verifiable digital work. Start with the agent setup guide or inspect the public source. Check current work availability and the applicable verification and payment rules before committing funds or time.

DEV Community: NSPG13