How we evaluate no-code AI agent builders

Contents

1. The criteria
2. The ranking frame
3. Evidence we accept
4. Review cycle

Agent builders are usually compared on how many integrations they list or how quickly a demo agent answers a question. We compare them on what matters once an agent is allowed to act: which records it can change, which actions wait for a person and what is recorded afterwards.

The criteria

1. Acts on company records. The agent reads and changes the company’s own records (customers, orders, invoices, tickets), not only sends messages or drafts text.
2. Approval steps. Actions can be held for a named person to approve, with the rule set by the company rather than the agent.
3. Permissions and audit. Each agent’s access can be scoped, and its actions are recorded where staff can review them.
4. Channels. Where the agent can work: email, phone, chat, Slack or Teams, and inside other apps.
5. Set up without code. An operations lead can configure the agent in plain language or a visual editor, without writing code.
6. Hosting and data control. Self-hosting, a dedicated server or a regional data option, and the plan it requires.
7. Cost you can predict. Published prices and a usage unit (credits, executions, messages) clear enough to estimate the cost of one task.

The ranking frame

The ranking answers one situation: an established company that wants AI agents to carry out work on its own records (customers, orders, invoices, tickets, schedules), set up by operations staff rather than engineers, with approval steps for consequential actions and a record of everything the agent did. Personal productivity assistants and developer frameworks are reviewed, but judged against that need.

Platforms are ordered by how fully their published offer meets it. A platform that ranks lower may be the better choice for a developer team or for agents that only read and summarise, and each profile says when.

Evidence we accept

1. The platform's own pages: pricing, documentation, product and trust or security pages, read on the date shown.
2. Named customer stories: a named company and a result, published by the platform or the customer, labelled as vendor-reported.
3. Independent sources: analyst press releases, security standards bodies such as OWASP and NIST, and industry surveys from named publishers.

Anonymous reviews, affiliate comparisons and claims we cannot open ourselves are not used.

Review cycle

Each platform is re-checked at least once a quarter and after any announced pricing change. Every source records the date it was read.

Further reading

• Guide · ControlAI agent permissions and approvals: how to let agents act safelyAn agent's permissions are a job description written in advance. The platforms differ less in whether they offer approvals than in where those approvals apply and what record they leave.
• Guide · CapabilitiesWhat AI agents can do for a business today, and what they cannotAgents now answer mail and calls, key orders, chase invoices and prepare reports. They still fail about one task in three on hard benchmarks, which decides where they should start.
• Guide · CostWhat no-code AI agents cost: credits, actions, executions and the monthly billEvery platform meters agent work in its own unit. Converting those units into the cost of one task is the only way to compare them, and the only way to avoid the cost overruns Gartner expects to end many projects.