Best AI agents for business, judged by an open-source test
Every list of best AI agents for business ranks its own author, so the name at the top names the publisher. Seven rows are what a buyer can check instead, applied to seven platforms. Kortix is the open-source AI Operating System, and the one that answers the self-host and approval rows in a file you own.
Seven platforms, seven rows, one test
The same questions asked of every vendor, Kortix held to them first.
The six competitor columns were read on their own sites in October 2026. A cell marked Not published is a question their pages leave unanswered.
Why every list of best AI agents for business ranks its author
Four of the pages ranking AI agents for business are written by vendors. Each puts its own product first. Gumloop's 8 best agentic AI tools gives slot one to Gumloop, "probably the best AI agent builder I've used". Dapta's list opens on Dapta, which "takes the top spot because it handles the job that turns directly into revenue". TinyCommand leads its summary with TinyAgents. Lindy's own 9 best agentic AI tools opens its comparison table with Lindy.
The number at the top measures the publisher's confidence in its own pricing page. A list shaped that way is a demo with a scoreboard attached.
The mechanisms underneath are still worth reading. Lindy's pricing page bills credits for every active person in a workspace. Zapier bills tasks. Intercom bills outcomes. Three different units of work, and they decide more about your year than any placement does. None of the six rivals publishes the test a buyer would write before delegating a write to an agent.
Five questions whose answers you can check on the vendor's own page in an afternoon, plus the two structural rows that make those answers checkable at all. Kortix is one of the vendors tested. The configuration behind its answers is shown in full as code.
The five questions that decide it
What it may write to. An agent that changes nothing is a chatbot with better manners. An agent that changes anything is a liability. Ask which connections it can act through, and whether that list is something a person can read before it runs.
Who approves the write. Every demo has an approval step. The question is where it lives and what it holds. On Kortix, session work reaches main through a change request. Merge is default-deny for agents.
Where the work happens. A shared runtime turns one agent's mistake into the next agent's starting state. One isolated sandbox per session. Each session has its own isolated machine and branch. Put the same question to any vendor and read how specific the answer gets.
What happens when the model changes. Vendors swap models, cap usage and reprice. The question is who holds that decision. On Kortix the buyer does: any model provider with your own keys, or the ChatGPT plan you already pay for.
What you pay per unit of work. Seats, credits, tasks and outcomes bill different things. Which unit your work lands in decides what a busy month costs. A seat bill stays flat while the work climbs.
Two structural rows make those five answerable: whether the code is open source, and whether you can run it yourself. An open-source agent is one whose access, memory and cost a buyer can inspect without asking the vendor for a favour.
Applying the test
Kortix answers the five questions out of one repository. Each agent's reach is a list in a text file you can diff. Sessions run isolated. Work lands through a change request, and the model is a key you hold. Both comments that carry that rule appear verbatim in the code panels.
Lindy bills every active person in the workspace, from $29.99 a month for 3,000 credits each (its pricing page, 27 October 2026). Those credits pool across the workspace. You pick a model for each task (checked October 2026). Gumloop sells a visual builder in its own cloud. Pro starts at $37 a month, and enterprise pricing is quoted separately. Zapier Agents act across 9,000+ apps through the connections you already have, and Zapier bills by task rather than by seat.
Microsoft Copilot Studio builds custom agents and workflows over 1,500+ connectors and MCP servers. Approval gates sit inside its application lifecycle pipelines, and audit logs route to Purview. Intercom's Fin answers support conversations at $0.99 per outcome. It charges nothing when it cannot answer, and counts one outcome per conversation. Salesforce Agentforce bills in Flex Credits or Conversations, or per person, with a PayGo option billed monthly in arrears.
Lindy answers the approval row too. Its pricing page says anything with outside impact waits for approval, such as sending an email, updating a ticket, posting to another channel or publishing a doc. Read-only lookups from approved sources go ahead. The other four say nothing about it. For them the vendor selling the agent is also the one deciding whether its writes get a second look. A file you own moves that decision to your side of the table. On the rows this test turns on, Kortix is the one to start with.
The two lines an agent's access rests on
From the starter manifest in the Kortix repository: the deny-by-default rule, and the grants the starter gives its own default agent.
# `agents:` map per project — GOVERNANCE ONLY (connectors/secrets/# kortix_permissions/skills/workspace/enabled), plus `file`, the path of the# agent's `.md`. Agent BEHAVIOR (description, model, mode, temperature,# permission, the prompt itself, …) lives entirely in that `.md`'skortix_version: 2
# v2 is deny-by-default: an omitted connectors/secrets/kortix_permissions/# skills grant resolves to "none", so a scoped specialist starts locked down.agents: kortix: file: agents/kortix.md connectors: all secrets: all
Buying questions, answered plainly
What is the best AI agent for business?
There is no single best one, because the answer depends on what the agent may write to and who approves that write. Kortix answers both in a configuration file you own, which is why it leads the comparison.
Which AI agent is best for a small business?
The one a small team can inspect. Pick a tool whose configuration it can read, whose spend follows usage rather than headcount, and whose setup it keeps if it stops paying.
Are AI agents worth the money?
They are when the work they finish would otherwise take a person an hour and you can see what they did. Credited, task-billed and outcome-billed work cost different amounts at the same volume, so ask which unit your month lands in.
Can AI agents be self-hosted?
Yes. An open-source agent runs on infrastructure you control, which is exactly what the self-host row measures. Hosted products in the table do not offer it.
What is the difference between an AI agent and a chatbot?
A chatbot returns text. An agent takes actions in your tools, which is why the write surface and the approval gate decide whether it can be trusted near production. A chatbot has nothing to approve.
Read the manifest before you trust the demo
Start a project and open kortix.yaml to see exactly what the agent is allowed to touch.
Open source · One isolated sandbox per session · Self-host is free