AI agents

An AI agent in Orvanta is a step type (module) you can drop into any flow, next to scripts, branches, and loops. Given a prompt and a set of tools, the model decides which tools to call, in what order, to produce a result — instead of you wiring that logic by hand.

What an AI agent step is

An AI agent step is configured with:

• Provider: which AI connection and model to use.
• System prompt / user message: the instructions and the task. The user message supports variable interpolation (e.g. flow_input.query).
• Output type: text (default) or image.
• Output schema: an optional JSON Schema the model’s response must conform to, for structured output instead of free text.
• Memory: an identifier that persists conversational state across separate invocations of the step, so the same agent can pick up context from a prior run.
• Attachments: images or PDFs passed in as S3 object references.
• Streaming: emits incremental events (token_delta, tool_call, tool_call_arguments, tool_execution, tool_result) instead of waiting for the full response.
• Temperature / max completion tokens: standard generation controls.
• Parallel tool calls: whether the agent can fire off multiple tool calls at once instead of one at a time.

Tools

The agent’s tools are what make it more than a chat completion. Each tool is one of:

• A flow module — any script or flow already in your workspace, wired up exactly like a normal step. The agent calls it like any other tool and gets its result back.
• An MCP tool — a reference to an external MCP server, with an optional whitelist (include_tools) or blacklist (exclude_tools) to limit which of the server’s tools the agent can see.
• A web search tool — for looking things up rather than acting on internal data.

Tool IDs and summaries can’t contain spaces (use underscores) — they’re passed to the model as function names.

Building multi-step agent patterns

There’s no separate “orchestrator” or “agent pool” primitive — an AI agent step is just a step, so you build agentic patterns the same way you build any other flow:

• Sequential handoff: chain AI agent steps like any other steps, wiring one step’s output into the next step’s input_transforms.
• Parallel work: use the flow engine’s parallel-branch construct to run independent steps (agent or otherwise) concurrently and merge their results.
• Human-in-the-loop: route an agent’s proposed action through a suspend/approval step before anything irreversible executes — the same mechanism used for any manual approval gate, not something specific to agents.

Related

• MCP: Connecting external tool servers to an agent.
• Approval steps: Gating an agent’s action behind human review.
• Workflows: How steps (including agent steps) are wired into a flow.
• Jobs and runs: Inspecting an agent step’s run trace, including tool calls.