Hey folks—wondering sth: chaining agents is possible if you wait for agent A to respond before asking agent B sth. Can it be a default behavior though in the instructions? I know it can be done via Make/Zapier/API but native solutions would have been useful. Example: "Explore accounts in Snowflake" then "build a pitch tailored for these accounts".
Thanks for your question Gilles Bertaux! You can do that when it's within a single agent. You can define multiple steps/tools to call in a sequence using the instructions - exactly like you mentioned. If you need to have different specialized agent for some reason, you'll need to rely on Make/Zapier to chain them.
Yes I feel the two agents might need separate specialisation, however it raises a broader question: what's the max amplitude of tools/instructions a single agent can have without losing relevance? Can a Snowflake explorer be able to do a Notion search and build a summary doc that match certain guidelines? (maybe not the best example, but you get the idea)
There is not hard rules. If it uses different tools and you don't need to add too many instructions to make them work separately, it can work well. Because the agent only "looks" at one at a time and then use output of all of them to generate the answer. So it usually works better than one could expect. It can be helpful to clarify in the instructions which tools it has access to and when and how to use which tools. 😊
Roger that, thanks for the clarification!