Agentic workflow tools judged on the machine, open-source Kortix first

What a canvas gives an agent, and what the run leaves behind

Comparison lists for agentic workflow tools count connectors, templates and model providers. Open the vendor documentation and a plainer difference appears. A workflow builder hands you a canvas: nodes joined by lines, each one a call to an app, a branch or a transform. An agent in that picture is a single node. It reads the payload the node before it passed along and returns something for the next one.

That design solves a real problem: making a few dozen SaaS APIs behave as one sequence. It also settles where the work happens, and vendors settle that differently. A Zap runs the steps you define, on Zapier's workers, calling endpoints your stored credentials already reach. Zapier Agents go further: you grant an agent a set of actions from Zapier's app library, and it chooses which to use and completes the fields each action requires (Zapier).

Gumloop answers the machine question another way. Each conversation runs in its own secure cloud sandbox, an isolated VM where the agent runs shell commands, creates, modifies and organises files, and installs packages with pip install and npm install.

So the machine is the first row to read, and the second is what the run leaves behind. Here each session gets its own isolated machine and branch, and session work reaches main through a change request. Merge is default-deny for agents. The definition covers the loop itself. The other comparison separates written steps from decided ones. The tools worth comparing are the ones buyers name for the job, and the row held to each is where its work stands when the run ends.

n8n, on its own terms

n8n is the builder most teams meet first, because it is the one you can run yourself. Its code is published under the Sustainable Use License, and files whose name or path contains .ee. need an Enterprise licence. It self-hosts, and pricing runs per monthly workflow execution. A team that wants a canvas, its own server and a bill tied to executions has a fair reason to choose it, once somebody has read the licence.

The agent in n8n still works as a node. When your process is a series of API calls with a model inside one of them, the graph is the right artifact. You version it, review it and keep it. The limit arrives later: what the agent may do is whatever that node type exposes, and the step belongs to the graph rather than to the work.

Zapier, on its own terms

Zapier trades depth for reach and asks nothing of your infrastructure. Task-based pricing runs across plans from a free tier, cloud only, and approval requests are listed as Beta. A small team connecting two SaaS products has no server and no wish to run one. For that team Zapier is the better pick, and the Beta label is honest about where the approval feature stands.

The run stays inside Zapier's cloud. You get the automation and a task count, and you can audit both. The agent works through the actions you grant it rather than a machine you hand it, and it picks an option for each required field itself.

The six rows to ask any vendor about

Every vendor fills the cheap rows. These six cost more to answer, and a reader can check each one against a page.

1. Where is the definition stored, and can you diff it? A file in a repository carries a history and a reviewer. A canvas inside someone's product carries a version you cannot export.
2. What does the run act on? A repository, a worker calling APIs, or a machine with its own filesystem and shell. Ask where the files sit once the run is over.
3. Who can use it, and at what level? One vendor answers with agent-scoped permissions on its Pro plan and org-wide permissions on Enterprise. Ask which level your team will sit on before the price matters.
4. Who reads the result before it lands? A change request, a pull request and a record edited in place are three different levels of review. Only the first two leave a diff.
5. What does one run cost at the cap? Per task, per execution and per credit fail differently once a busy month arrives. Ask for the ceiling.
6. Can you self-host it? An offer you can test this week settles more than a roadmap line. It also tells you who holds the machine.

The two hard rows, answered in files

Both rows come down to what a file declares. kortix.yaml names the sandbox image, with its CPU, memory and disk. Each session boots its own cloud computer and cuts a branch named after it. Session work reaches main through a change request, and merge is default-deny for agents.