Split an assistant's work across agents, so questions go to one that answers in text and build requests go to one that produces apps. Agents are added in the editor of an assistant you have already created in Make an assistant.
How routing works
When a message arrives, a router selects one agent to handle it, and that agent alone writes the reply. The router reads the opening of each agent's instructions and matches the request against them, so what an agent says about itself determines the work it receives.
one message, one agent
"what can I use instead of buttermilk?"
↓
router
↓ reads the opening of each agent's instructions
Recipe Recall
specialist
← selected, writes the reply
Planner Builder
toolchain
The other agent is not consulted.
1
Add an agent
Open an assistant from Your assistants, choose Edit, and use Add Agent. Enter an Agent Name and its Personality & Goals.
Your assistants
→
Edit
→
Add Agent
Use a plain description of the job as the name, such as "Recipe Recall" or "Planner Builder", since the name appears in the sidebar and in routing.
2
Choose the type
agent types
Type
What it produces
Extras
Specialist
A text reply
Instructions, Memory, Files
Toolchain
An interactive app
Adds Apps and memory packs
Choose specialist for questions, lookups, summaries, and conversation. Choose toolchain to build apps, which is the only type that can, so an assistant that only replies in text does not have one yet. The type is fixed once the agent is created.
3
Write the instructions
Open the Instructions panel under the agent and describe the kinds of request it handles in concrete terms. The router reads roughly the first 600 characters, so put that description at the start.
the opening of an agent's instructions
Answers questions about the recipes in memory: substitutions, cook times, what pairs with what, and how to scale a dish up or down.
Why this matters
Instructions addressed to the router, such as "this is the default agent", do not affect routing. The router matches the request against the description rather than taking orders from it, so naming the work moves routing where naming the priority does not.
4
Verify which agent answers
Press Save, then use Open chat and send the questions you expect people to ask. Send one intended for the specialist and one build request intended for the toolchain agent; each reply comes from a single agent, which shows whether routing lands where you intended.
When a message reaches the wrong agent, edit the opening lines of both agents so each names its own kinds of request more specifically. Sharpen the descriptions rather than adding rules about precedence.
Give an agent its own memory
Add notes under the agent's Memory section, which works the same way as the assistant's but is loaded only when that agent handles a message. Assistant-level notes are carried by every agent. Adding notes is covered in Add memory to an assistant.
Turn on memory packs
Switch on a memory pack under a toolchain agent to add its examples to that agent's memory. The packs cover Gaming, Productivity, Creativity, Lifestyle, and Business, and each one makes the builder better within that category while increasing token usage on every build.
Rename and delete agents
Rename an agent, rewrite its instructions, or delete it from its panel in the editor. Deleting an agent removes it from routing. Apps that were already saved keep working, since a saved app runs its stored program and does not call back into the agent that wrote it.