Where the run happens, and who reviews it
Whose computer the run works on, and who reads what comes back: those two questions decide whether an agentic workflow is worth keeping. Every run in Kortix, the open-source AI Operating System, boots its own cloud computer, and the work a session commits reaches main through a change request, so both answers are yours to change.
The open-source AI Operating System
One isolated sandbox per session · Self-host is free · Run it in your VPC or on your own on-prem network
main through a change request. Merge is default-deny for agents.triggers: - slug: daily-digest type: cron agent: kortix cron: "0 0 9 * * 1-5" timezone: America/Los_AngelesSource: kortix.com/automations
Each card is a trigger row from kortix.yaml and the work its run leaves behind.
At 09:00 on weekdays the digest session reads yesterday's commits and leaves the summary as a change someone can read.
Fires weekdays at 09:00 in America/Los_Angeles, one fresh session per run.
The payload renders into the prompt, the session starts on a computer of its own, and the CRM entry waits as a change.
A webhook trigger without its signing secret is rejected at validation.
The support agent starts on its own computer and leaves the handoff note on its branch for the next shift to read.
A triggered session is visible to the whole project, not private to whoever set the trigger.
Open source means the workflow is a file you can read
More tools, judged the same way, are on the tools comparison.
kortix.yaml, in the repo you owngh aw compile pins into an Actions workflowmainSource · KortixSource · GitHub Agentic Workflows
Gumloop's sandbox docs (https://docs.gumloop.com/core-concepts/agent_sandbox_and_secrets) cover the isolated VM each conversation runs in and what the agent may do inside it, its Human in the Loop docs (https://docs.gumloop.com/core-concepts/human_in_the_loop) cover the approval a person gives before a gated tool call, and its pricing page (https://www.gumloop.com/pricing) lists plans, permission levels and no self-host offer. GitHub's Agentic Workflows docs (https://github.github.com/gh-aw/) cover the repository a run touches and the review before a merge, and its licence (https://github.com/github/gh-aw/blob/main/LICENSE) is MIT.
An agentic workflow is a job an AI agent finishes on its own, from a trigger to a result you can read. The run plans the steps, uses tools on a computer, and keeps going when one step fails. The full definition.
Not by default. Nothing pauses the work unless you set a gate, so the approval has to sit where the work lands. In Kortix, session work reaches main through a change request, and merge is default-deny for agents.
An automation runs steps you defined in advance. An agentic workflow gives the agent a goal and a computer, and it works out the steps itself, including what to do when one fails. The trigger can look the same.
Yes. Kortix is the open-source AI Operating System, and a trigger there is a row in kortix.yaml beside the prompt it fires. Self-host is free.
On a computer the agent controls. Each Kortix session gets one isolated sandbox and its own branch, so the run works on a real Linux machine and leaves a diff.
A weekday digest at 09:00 that reads yesterday's commits and opens a change request with the summary. The trigger is a row in kortix.yaml, the run gets its own computer, and a person reviews the change. Build one.
One repo holds the agents, the triggers and the memory, and self-host is free.