What Matters
An AI-agent framework, a hosted business assistant and a funded trading agent are different products. Choose by the task, the permissions required and the evidence available. This directory compares eight documented options and shows how to design a small pilot. It does not rank returns, certify security or claim to cover every agent.
Public-source review: . Original publication: 16 May 2026.
DN Evidence Block: AI-assisted desk review of the official documentation linked below. Scope: eight platform families, their stated purpose and important operating constraints. No accounts created, software installed, agents executed, funds deposited, trades placed or withdrawals tested for this review. No independent reviewer is recorded. Public documentation confirms what a provider describes, not that a particular deployment works or is safe.
Commercial disclosure: This directory uses direct documentation links, with no referral registration links. Other DN pages linked here may contain disclosed affiliate relationships. Compensation is not evidence of product quality; this page does not select a trading venue or recommend buying an agent token.
Browse the directory · Fetch.ai and the free tier · Olas and Pearl · Build your pilot checklist · Estimate operating cost
DN Alpha Thesis: the permission boundary matters more than the agent label
Our assessment: start by asking what the system is allowed to change. A research assistant that drafts a report has a different failure cost from an agent that sends payments, trades assets or changes customer records. A promising demonstration should lead to a bounded test, not broader permissions by default.
For this guide, an agent can choose actions or tools toward a task; a workflow follows defined steps and may contain agent steps. A framework helps a developer build that system. A hosted service runs some of it for the customer. A token associated with a project is neither the software nor proof of useful execution.
AI-agent directory by job and operating model
Selection is deliberately limited to eight options with accessible first-party descriptions. The rows are grouped by use case, not ordered by quality. Evidence status for every row: public documentation reviewed; end-to-end operation untested by DN. Country, account, plan and feature restrictions must be checked separately.
| Platform and official source | Documented role | DN pilot boundary |
|---|---|---|
| Fetch.ai: uAgents and Agentverse | Build or connect agents; hosted and external deployment paths, discovery and interaction analytics. | Start with a harmless message or public-data response. Check runtime, subscription and quotas; a marketplace listing is not a safety audit. |
| Olas and Pearl | Agent infrastructure and an application for operating agents, including prediction-market agents. | Check the exact agent's current availability, market eligibility and funding needs. Separate trading outcomes from token incentives. |
| Coinbase AgentKit | Developer toolkit connecting agents to wallets and onchain actions; Python and TypeScript packages. | Use testnet first. The toolkit alone does not enforce transfer approval, spending limits or destination allowlists. |
| LangGraph | Low-level orchestration for stateful workflows, persistence and human-in-the-loop steps. | Test interruption and restart with synthetic data. A persistence feature does not prove your workflow avoids duplicate external actions. |
| CrewAI | Flows manage state and execution; Crews organize agents collaborating on delegated tasks. | Begin with one bounded task. Measure whether adding agents improves accepted results after model cost and review time. |
| n8n | Workflows combining model providers, tools, memory and AI components. | Use a test workflow with read-only credentials. Confirm hosting, licence and connector terms for your intended use. |
| Microsoft Copilot Studio | Graphical agent and workflow creation with organizational data and connectors. | Check tenant permissions and licensing. Pilot with a small approved knowledge set before enabling record changes. |
| GitHub Copilot cloud agent | Repository research, code changes and tests in a cloud development environment, with reviewable changes. | Use an isolated branch and require review and tests before merge. Generated code and session logs are not independent verification. |
These tools can be combined, but they are not interchangeable. For example, an orchestration framework does not by itself supply a lawful market account, a trading strategy or a secure signing policy. A business agent connector does not establish permission to disclose the connected records.
How do Fetch.ai uAgents and Agentverse differ, and is there a free tier?
uAgents is the Python framework; Agentverse provides hosted-agent and external-agent discovery paths. The current local-agent guide describes running a uAgent on your own machine or infrastructure and implementing the Agent Chat Protocol to connect it to Agentverse and ASI:One. Local execution still needs an available machine and any services your code calls.
The subscription and quota documentation explicitly describes a free tier. It says users must log in at least once every 30 days to keep their mailbox active. It also distinguishes plan access from resource quotas. Do not interpret the page's example quota figures as guaranteed allowances for your account; inspect your current plan and usage dashboard.
- Choose a local or hosted path based on who will operate the runtime.
- Follow the current official guide with a non-financial task and non-sensitive data.
- Check messages, errors and quota behaviour before connecting additional tools.
- Budget for model calls, hosting and third-party APIs separately. A free platform tier does not make those inputs free.
Agentverse's verification guide distinguishes domain ownership from other verification types. A domain badge is an identity signal, not evidence of investment performance or a security guarantee. Use the current documentation rather than assuming a tutorial labelled 2025 still reflects today's interfaces or limits.
Olas, Autonolas and Pearl: availability is specific to the agent
Olas is the ecosystem also known as Autonolas; Pearl is an application for operating agents in it. They are separate from Fetch.ai's uAgents and Agentverse. Using one does not imply access to the other.
The official 5 May 2026 Pearl update paused new Optimus, Agents.fun and Polystrat instances. That is not the latest word on Polystrat: the 1 July Q2 update says Polystrat was adapted to Polymarket's wallet changes and new agents could be created again. Do not extend that reopening to every agent previously paused.
The 9 October activity update describes configurable targets for Polystrat, Omenstrat and Pearl Connect, and warns that more activity uses more balance on fees and gas. This supports a dated account of the product's evolution; it is not a DN-funded test of current onboarding or execution.
Before operating a financial agent, verify your eligibility for the underlying market, custody and signing arrangements, all running costs and how to stop it and recover funds. Prediction markets can have gambling or financial-product restrictions depending on jurisdiction. Do not bypass location, identity or account restrictions.
The Olas prediction dashboard separates trading return from total return including staking rewards and reports different observation windows. A selected winning position or incentive-supported return does not demonstrate repeatable net trading profit. DN does not validate those provider-reported metrics in this directory.
Wallet-enabled agents need controls outside the prompt
The AgentKit risk section warns that third-party text can influence an agent's actions and that the SDK does not itself gate transfers behind human approval, cap spending or restrict destinations. This is an especially important distinction between an integration toolkit and a complete controlled application.
DN's pilot rule: do not connect a funded wallet or production signing authority until approval, amount, destination and revocation controls have been implemented and tested at the action boundary. A sentence telling a model to be careful is not an enforceable spending policy. For documented test scopes and limitations, see the DN Agent Trust evidence hub; its results apply only to the systems and versions described there.
DN Agent Pilot Checklist
Open each step and record the answer in your own notes. This worksheet does not collect or save answers, run an agent, calculate a safety score or approve a deployment. It works without JavaScript. An unmet permission, eligibility or recovery condition stops the pilot; unknown means unresolved.
1. Define one task and an acceptance test
Write down the input, expected output and who decides it is correct. Example: summarize ten public documents with a source for each factual claim. Keep the same task set for the manual baseline and agent trial. Do not count fluent output as a completed task.
2. Set the permission and data boundary
List every tool and account the agent can reach. Start read-only with synthetic or public data. Name the actions requiring prior approval, especially payments, trades, deletion, publishing and external messages. If you cannot restrict access to what the task needs, stop.
3. Budget the whole run
Include model and API charges, hosting, connector or subscription costs, human review and failure recovery. Set a maximum cost and run duration. For financial agents, account for gas, trading fees, funding and capital at risk separately; a software budget is not a loss limit.
4. Test failure, stop and recovery
In an isolated environment, test a missing dependency, an expired permission and a repeated request. Confirm how the operator stops execution and revokes access. Check that retries do not duplicate a payment, message or record change. Keep credentials and personal data out of test reports.
5. Compare outcomes before expanding access
Record attempted tasks, accepted tasks, corrections, total cost and operator time. Compare with your manual baseline over the same task set. A small successful pilot supports that limited use case only. Failed controls or unresolved errors mean repair and retest, not automatic rollout.
What does an agent cost per accepted task?
Cost per accepted task = total pilot cost ÷ tasks meeting the agreed acceptance test. Count rejected and failed attempts in total cost. If no tasks pass, the cost per accepted task is undefined; report the spend and zero accepted tasks instead of a zero-cost result.
Fictional example: 100 attempts consume $20 of model/API usage and $10 of allocated hosting. Review and corrections take two hours, valued at $25 per hour. Total pilot cost is $80. If 40 tasks meet the acceptance test, cost per accepted task is $2.00. Dividing $80 by all 100 attempts gives $0.80 per attempt, a different measure. These figures are not provider prices or a forecast.
For a trading experiment, a completed order is not automatically a successful investment. Evaluate fees, risk and outcomes separately. The DN trading automation comparison helps distinguish operating approaches; it does not establish profitability or confirm eligibility for a particular venue.
What this directory deliberately does not claim
- No universal winner: these products serve different jobs and have not been tested on a common benchmark here.
- No blanket production readiness: public documentation, code availability and a functioning landing page do not validate an integrated deployment.
- No token recommendation: software adoption, token price and tokenholder rights are separate questions.
- No regulated-use certification: this review does not establish suitability for clinical, legal, credit or investment decisions. Those require use-case-specific evidence, professional oversight and applicable legal review.
- No closure inference: a platform omitted from this selected directory has not thereby been classified as inactive, unsafe or inferior.
Choose your next step
For a business workflow, complete the checklist before granting account access. For an agent that moves money, start with the versioned Agent Trust evidence. For further practical resources, browse DN Agentic Tools. Use the source links in the directory to check the exact service, version and current terms before committing time or money.
Sources, limitations and correction record
Sources are the official documentation, repository and dated product updates linked beside the relevant claims. The review date records when those materials were checked. Availability, quotas and capabilities can change. We did not conduct account-level checks, independent performance benchmarking or a security audit.
27 October 2026 correction: DN withdraws the earlier claims of exhaustive coverage, universal production deployment, unsupported rankings, selected-return endorsements and unverified pricing assertions. The article is now a selected, source-linked directory with a manual pilot worksheet. Historical examples removed from this version should not be treated as current recommendations. The original URL and publication date are retained.
Report a specific error through DN's contact page, ideally with the claim, current official source and date. This is general educational information, not personalized financial, investment, legal or tax advice. Automation can cause data loss, unauthorized actions and financial loss. Crypto assets and leveraged or prediction-market positions can lose all committed funds.
