Skip to main content
Runchat is a tool building platform. Anything you publish as a Tool can be used automatically by language models in the chat agent or the Agent node. This gives the agent the ability to generate and edit images, create meshes, search the internet, research concepts, evaluate designs and much more.

​
How do agents use tools?

When you (or someone else) publishes a workflow as a tool, you provide a description of what the tool is supposed to do as well as names, types and descriptions for all input parameters. This information is given to the agent along with your prompt, and the agent can decide whether or not to use a tool to complete your request. The agent formats its response to match that expected by the tool, and Runchat then runs the workflow with the inputs provided by the agent. Once the workflow completes, the outputs are sent back to the agent so that it can decide what it wants to do next. This is called a tool calling loop, and for complex queries this might happen dozens of times.

​
Adding tools to the agent

Tools attach in two places, depending on whether you’re working with chat or a node:
  • Chat: The chat agent can search your own private published tools as well as the database of publicly available tools and then inspect and run these automatically.
  • Agent node: in the node’s settings bar, click Select Tools to attach tools to the node. The Tool Effort setting controls how thoroughly the agent uses them (Fast, Moderate, Max, Custom).

​
I want to connect Runchat to some service (e.g. discord or slack), is there a tool for this?

You can get the Runchat agent to help create custom tools for you. Click in the chat input at the bottom of the editor and describe what you want to do. Don’t worry if the agent doesn’t get it right first try. You can:
  • ask the agent to fix errors
  • run the node and then ask it to view the output to improve the node
  • use a smarter coding model (usually higher credit cost = smarter model)
Sometimes it helps to include a link to relevant API documentation, and add a search tool so that the agent can read the docs and implement the correct specification. If you make a useful tool, share it on Discord so others can use it.

​
An example: Design Product Research

Research is a task that requires lots of trial and error. People often look for things, learn something then circle back to explore some other avenue of investigation or idea. As a result, research is not particularly well suited to a linear workflow because it requires loops. This is what the Agent is perfect for. Lets say you want to research some tiles for a new design project. For this task, let’s use the chat agent:
  1. (optional)add some pre-existing ideas or imagery you might be interested in to the canvas.
  2. In chat, say “we’re researching award winning residential architects working with ceramics. Can you find some exemplary projects and produce a short summary with images?”
If we run this prompt, the Agent will find a search tool to experiment with different queries to try and find something that matches our request. It might try searching for ceramic architecture first, or award winning ceramic, or one then the other. Its non-deterministic. After it’s found some projects, it might start searching for individual architects. And after that, it might start searching for specific products. We can control this behaviour by being more or less specific with our initial prompt.

​
Planning and Sub-Agents

You can use the agent to help construct an initial plan for a task then execute on it, or just go back and forth to dive deeper into areas you want to explore. You can also create workflows that contain the agent node pre-set with prompts and tools, then use another agent node to call these. This allows you to effectively delegate tasks to many sub-agents.

​
Using design tools

You can use sub-agents like this to call more powerful multi-modal models only to evaluate images, or to call more powerful models to write snippets of code, or to call many fast, cheap models to parse website content in parallel and so on. You can also use agents to call your design tools. Try using the agent node to run multiple workflows instead of building them with the node editor — you might be surprised what results!