What is the difference between an agent and a workflow?

Anthropic describes workflows as models and tools following predefined code paths, while agents let the model direct its own process and tool use. Many products sold as agents combine both.

How do I estimate what an agent will cost to run?

Ask for the input and output tokens used in a typical run, multiply by the model's per-million-token prices, and add any tool fees. Then multiply by the number of runs you expect each month.

Should an agent ever act without approval?

Low-risk, reversible actions can run automatically. Payments, deletions, outbound email and public posts should wait for a person to approve them.

Does GreenGo test how an agent behaves?

No. Green Light runs automated checks on DNS ownership, security headers, exposed secrets, known vulnerable dependencies and delivery files, and it is not a security audit. Sellers can also order a human review for $149.

AI agents you can buy today: a buyer's guide

Short answer: Buying an AI agent usually means one of three things: code you run with your own model API key, a hosted service you subscribe to, or a pay-per-task service. Before you pay, find out what each run will cost in model fees, what data and permissions the agent needs, whether it keeps logs, where a human approves actions, and what the license and update terms are.

What "buying an agent" actually means

Anthropic draws a useful line between workflows, where models and tools follow predefined code paths, and agents, where the model directs its own process and decides which tools to use. Most products sold as "agents" sit somewhere between the two. What matters more for a buyer is how the agent is delivered.

Delivery type	What you receive	Who pays model costs	Where your data goes	You control updates?
Code you run	Source code or a package	You, on your own API key	Your servers and your model provider	Yes
Hosted service	Login or API access	Usually included in the price	The seller's infrastructure	No
Pay-per-task	Results per job	Built into the per-task price	The seller's infrastructure	No

Code you run gives you the most control and the most responsibility. Hosted and pay-per-task options are faster to start but ask you to trust someone else with your data.

The buyer's checklist

1. Model and API costs

If you run the agent yourself, you pay the model provider for every token. For example, Anthropic's published API prices, per million tokens (MTok), include:

Model	Input	Output
Claude Sonnet 5.5	$2 / MTok	$10 / MTok
Claude Opus 5.5	$4 / MTok	$20 / MTok
Claude Haiku 5.5	from $0.10 / MTok	from $0.50 / MTok

Here's how to turn that into a cost per run. If one run uses 40,000 input tokens and 4,000 output tokens on Sonnet 5.5, that's $0.08 for input plus $0.04 for output, or $0.12 a run. At 500 runs a month, that's about $60 before any tool fees. Tools add their own charges: Anthropic's web search tool, for example, costs $10 per 1,000 searches. The Batch API takes 50% off for work that doesn't need an instant answer.

Agents loop, so costs can climb. Anthropic notes that agents trade latency and cost for better task performance, and that their autonomy means "higher costs, and the potential for compounding errors." Ask the seller for a measured cost per typical run.

2. Data access and permissions

OWASP lists "Excessive Agency" in its 2025 Top 10 for LLM applications. It comes from three things: too much functionality, too many permissions and too much autonomy. Ask:

Which tools can the agent call, and what can each one change?
Can it work with read-only access?
Does it need your admin or owner keys, or can you create a limited key just for it?
If it uses MCP servers, which ones, and with what scopes? (See What is an MCP server?)

3. Logs

You should be able to see what the agent did, step by step: which tools it called, with what inputs, and what came back. OWASP notes that logging and monitoring don't prevent problems, but they limit the damage and make them easier to trace.

4. Human in the loop

High-impact actions, such as sending email, moving money, deleting records or posting publicly, should wait for a person to approve them. OWASP recommends human approval for these actions, and enforcing permissions in the systems the agent touches rather than trusting the model to decide what's allowed.

5. License

For code, check whether you can use it commercially, change it, deploy it for clients and run it on more than one project. For hosted services, read the data-use terms: can the seller use your inputs to train or improve anything?

6. Updates

Models and APIs change often. Ask how long updates last, how the seller handles model deprecations, and whether there's a changelog.

Red flags

It asks for full-access or wildcard permissions instead of the minimum it needs.
No list of the tools it can call or the actions it can take.
The seller can't tell you roughly what a run costs.
No logs, or logs you can't export.
It takes irreversible actions (payments, deletes, outbound email) with no approval step.
Real API keys or .env files inside the delivery files.
Promises of specific income or results.
Obfuscated code in a product sold as "source code."

Threats to know about

The OWASP Top 10 for Agentic Applications, published December 9, 2025, names the risks that are specific to agents. A few that matter most when you're buying:

Agent goal hijack: instructions hidden in emails, web pages or documents redirect the agent.
Tool misuse and exploitation: the agent uses a legitimate tool in a harmful way.
Identity and privilege abuse: the agent acts with more access than the task needs.
Agentic supply chain vulnerabilities: a plugin, MCP server or dependency is compromised.

You don't need to be an expert to use this list. Ask the seller how their agent handles each one, and judge the clarity of the answer. NIST's Generative AI Profile (NIST AI 600-1) is a free, deeper reference if you're buying for a company.

How to compare listings on GreenGo

Browse the agents catalog, or search everything in Explore. When you compare listings, line up the same details for each:

Delivery type: code, hosted or pay-per-task.
Model required and an estimated cost per run.
Tools and permissions the agent uses, including any MCP servers.
Approval steps for high-impact actions.
License and update terms.
Green Light status and date. Green Light runs automated checks on DNS ownership, security headers, exposed secrets, known vulnerable dependencies and delivery files. A pass is shown for 90 days with the check date. It is not a security audit or certification, so you still need the checklist above.

Paid purchases on GreenGo include 14-day Buyer Protection, and buyers pay no extra fees on top of the listed price.

Sources

Updated October 20, 2026.