IVY AGENT
A workflow where every step meets the gate
Nodes, model turns and branches in one file. Describe it in a sentence, or build it on a canvas.
An ivy-agent is a graph of steps. Most steps are ivy-nodes, some are a model turn, some are a branch or a loop. IVYX registers the whole graph as one capability, so it is called like anything else in the workspace, and each step inside it is still its own call through the gate.
Support Ticket Triage
Analyze support tickets to determine team and urgency, then email urgent ones to on-call.
- 01extract the ticket text from the input messageModel turn
- 02Jev · ChoiceIvy Node
- 03Jev · ScoreIvy Node
- 04only when urgency is top levelBranch
- 05Email SendIvy Node
ivy.agent.ticket-priority-router, as published: the steps in the order its file lists them. See this agent
Three kinds of step
An ivy-node
A Python capability with a typed contract. The agent calls it through the gateway, the way any other caller would.
A model turn
One bounded exchange with the model your workspace is set to use. It can be given tools, and each tool call it makes is checked on its own.
A branch or a loop
A branch takes a path when a value is true. A loop runs a step once for each item of a list.
From a sentence to a file
Describe the job and IVYX writes the agent as a file in your workspace and opens it on a canvas. The same works from a terminal, with ivyx agent generate. Or start from an empty canvas and add the steps yourself.
What comes out is a draft you can read, not a finished product. Of the agents we publish, some were corrected by hand after a real run, and each one carries the test cases it had to pass.
One capability outside, many calls inside
It runs as a capability
The agent is registered under its own id, such as ivy.agent.sales-report. The Run button, an agent in chat and an MCP client all reach it through the gateway.
The gate reads the whole graph
An agent reaches the gate as process.exec, plus model.invoke when it has a model turn, plus what its nodes declare. A rule about file writes holds an agent that writes a file.
A turn starts narrow
A model turn stops at eight steps unless the file says otherwise, sees no MCP server unless the step names it, and its tools can be narrowed by risk, side effect, tag or origin.
A held step waits for you
When policy stops a step for a person, the run waits and the question appears as a card. Your answer resumes that step.
Beyond the Run button
Install it with its nodes
Adding a published agent from the Ivy Marketplace fetches every node it needs first, and writes nothing if one is missing.
Serve it as a tool
Deploy adds the agent to the set your workspace's MCP server offers, so Claude Code or another MCP client calls it by name, as an agent, under your policy.
Run it on a schedule
ivyx flow schedule writes a line in the project and a timer in the operating system. When it fires, the run is handed to a Studio window open on the project.
What it does not do
Not a sandbox
An agent runs its nodes on your workspace's Python kernel. The gate decides whether a step may run. It does not contain the step once it runs.
It needs a Python server
Steps that are nodes need one running. Without it the run is refused, with a sentence that says so.
Some agents need a key
Some published agents call an outside service and need your own API key for it. Each agent's page lists what it needs.
Verified means its own tests
A verified agent is one whose own test cases passed on its publisher's machine. The registry checks that record and does not run the cases again.
A published version stays
A version can be published once and cannot be withdrawn. Publishing and installing need the desktop app.
Run history shows the ends
The run history of an agent shows what went in and what came out, not yet the steps in between.