Tool Chaining in Single-Agent Systems

[sonnylaskar]

Hi All,

I have a use case involving a single-agent system with two tools. The goal is for the system to:

1. Receive a user question.
2. Call Tool 1 to process the question.
3. Automatically use the output from Tool 1 to trigger Tool 2 without additional instructions from the user.
4. Return the final result to the user.

Currently, I can achieve this behaviour by embedding the chaining logic directly into the system prompt of the assistant. However, I would like to explore a solution where this process is handled entirely through the definitions of the tools themselves, rather than relying on explicit instructions in the system prompt.

When I attempted this approach, the agent would call Tool 1 and provide its response. If the user subsequently instructs the system to execute Tool 2, it works, but I would prefer to eliminate the need for the user to explicitly trigger the second tool. Ideally, the system should seamlessly transition from Tool 1 to Tool 2 without additional user input, without requiring complex multi-agent setups or explicit chaining instructions in the system prompt.

Could you please advise on how to achieve this?

Thanks in advance!

[ekzhu]

Can you combine tool 1 and tool 2 into a single tool? If there isn't any model inference needed?

If you want to customize how tools are being managed it's better to implement your custom agent. We can eventually add some tool chain abstraction into the main agent.