Skip to main content

Updated Nov 1, 2026Verified against the product Oct 31, 2026

LLM steps (AGENT)

An AGENT step runs an LLM with a system prompt, a user prompt, optional tools, and optional multi-turn behavior. Use it whenever you need summarization, classification, routing decisions, or natural-language output—while keeping deterministic work in CALL steps.

Build with AI

Prompt-heavy agents are easiest to iterate via MCP; see Vibe coding.

Model sizes​

model_typeWhen to use it
smallStructured extraction, tagging, short summaries, JSON you will parse downstream.
largeMulti-input reasoning, long narratives, nuanced recommendations.

CALL steps vs tools on LLM steps​

FunnelStory has a shared catalog of functions — semantic.query, slack.send_message, email.send, tasks.create, CRM read/update, dataset operations, web search, and more. The difference is who decides when to call them:

  • CALL steps let you (the flow author) invoke any function in the catalog with arguments you control. You decide which step runs the function and what values go in.
  • Tools on an AGENT step let the model invoke functions during the step. Most of the shared functions can be attached as agent tools; a few are CALL-only.

Functions not available as agent tools remain usable in ordinary CALL steps. See Functions reference for the full catalog.

Available agent tools​

An AGENT step can be given most of FunnelStory's flow functions as tools — the same catalog CALL steps use (a few functions are CALL-only). Common choices include:

  • semantic.query
  • email.send, slack.send_message
  • tasks.create
  • CRM reads and updates (hubspot.*, salesforce.*)
  • drive.*, enrich.*, search.web
  • document builders such as pptx.create and pdf.create

See the Functions reference for the full set.

Each tool entry maps a name (what the model sees) to a function_id (what actually runs):

{
"tools": [
{
"name": "query_workspace",
"function_id": "semantic.query"
}
]
}

Fixed arguments​

For each tool, you can lock specific arguments so the model cannot choose them. Fixed keys are merged at runtime and hidden from the model's view of the tool schema — the model only sees the remaining parameters it is allowed to fill in.

Use fixed arguments to narrow scope: force a specific connection, pin a Slack channel, or constrain part of a query.

{
"tools": [
{
"name": "post_to_slack",
"function_id": "slack.send_message",
"fixed_args": {
"connection_id": "slack_conn_123",
"channel_id": "C01234567"
}
}
]
}

In this example the model can only choose the message content; connection and channel are locked by the flow author.

Prompting for JSON​

When downstream steps parse JSON, add explicit instructions — for example: "Output ONLY raw JSON, no markdown, no backticks." to the system prompt.

Threads, memory, and multi-turn (advanced)​

These are optional add-ons for specialized flows. Most agents do not need them.

FieldPurpose
thread_idOptional stable id so repeated runs can restore prior LLM context.
variable_storeWhen true, exposes persistent key-value storage tools (variable_get, variable_set, variable_push, variable_pop, variable_clear) for scratchpad memory inside the step.
multi_turnWhen true, allows the agent to pause execution and wait for external input before continuing (uses the signal tool path).

Details on these capabilities belong in the flow authoring guide.

Chat responses​

For agents you talk to in the product Chat experience, the final AGENT step should write its user-visible text to the global @.response variable ("out": { "set": "@.response" }). If you skip this, the thread may complete without showing a reply.

Example​

{
"id": "analyze",
"op": "AGENT",
"next": "",
"out": { "set": "@.analysis" },
"agent": {
"model_type": "small",
"system": "You are an analyst. Output ONLY raw JSON, no markdown, no backticks.",
"user": "Classify renewal risk for this payload:\n{{ $.account_payload }}",
"tools": [
{ "name": "query_semantic_db", "function_id": "semantic.query" }
]
}
}