Griptape Create: Agent from Config
The Config-First Route
- config
- key_value_replacement
- AGENT
Griptape Create: Agent from Config is the sibling of the main Create Agent node with the prompt stripped out. It builds a Griptape agent purely from a configuration node - the CONFIG input that decides which LLM, embedding model, and vector store your agent uses - and hands you back just the agent object. No prompt, no chat, no reply. Just a configured brain you can pass to task nodes that do the talking later.
What it's for
The main Create Agent node is all-in-one: configure, prompt, answer. This one exists for when you want the configuration as a reusable asset. You build an agent once from a CONFIG, then hand that AGENT to multiple task nodes - prompt tasks, image tasks, tool tasks - and each one runs against the same configured brain without you re-specifying drivers every time.
It also pairs with the pack's "default agent" machinery: the README covers setting default agent configurations, and a config-built agent is the natural thing to promote to a default. If you find yourself wiring the same config into five nodes, this node is the DRY approach.
How it works
It's a thin wrapper over the same gtComfyAgent the main node uses, but it only reads the config:
Defaults.drivers_config = config
create_dict["drivers_config"] = Defaults.drivers_config
self.agent = gtComfyAgent(**create_dict)
No prompt is constructed and nothing runs. It just instantiates the agent with the drivers the config node describes, which is why the only outputs are the agent itself.
Inputs
config(CONFIG) - the driver configuration. If omitted, the default configuration is used (which means whatever default config the pack has - often requiring an API key).key_value_replacement(DICT) - still available here for context placeholders the agent will use later.max_subtasks(INT, default 3) - the subtask cap, carried over for when this agent eventually uses tools.
Notice what's not here: no STRING, no tools, no rulesets, no agent input. The node deliberately drops them - its whole point is configuration-only creation. If you need tools or rules, create them on the task nodes you feed this agent into, or use the full Create Agent node instead.
Output
AGENT- the configured agent, ready to wire into task nodes.
Installation
Ships in ComfyUI Griptape Nodes:
- ComfyUI Manager: search "Griptape" → install.
- Manual:
cd ComfyUI/custom_nodes && git clone https://github.com/griptape-ai/ComfyUI-Griptape, restart ComfyUI.
Gotchas
- Since there's no STRING input, you can't get a text answer out of this node. If you wire it up and see no OUTPUT, that's correct - feed it to a task node that returns text.
- The
configyou attach has to actually specify a model. The pack checks for a blank model and raises a clear error; a config node with nothing selected will trip it. - Use it when you want one configured agent shared across many nodes. For a one-off "configure and chat in one place," the full Create Agent is simpler.
This is the "build the brain once, use it everywhere" node - less flashy than its sibling, but the one that keeps big workflows from becoming configuration spaghetti.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| configopt | CONFIG | The configuration for the agent. If not provided, the default configuration will be used. | |
| key_value_replacementopt | DICT | The will replace the {{ key }} with a value. | |
| max_subtasksopt | INT | 3 | The maximum number of subtasks to run when an Agent uses a Tool |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| AGENT | AGENT | Text response from the Agent |