Skip to content

README.md

Latest commit

 

History

History
59 lines (43 loc) · 2.3 KB

README.md

File metadata and controls

59 lines (43 loc) · 2.3 KB
Raw

AgentNode LLMAgent External Tool Resume

This example demonstrates a GraphAgent workflow where an AgentNode runs an LLMAgent with node-scoped external tools.

The child LLMAgent emits the external tool call. A following normal graph node performs the checkpoint interrupt and later feeds the caller-provided tool result back to the same AgentNode.

The flow is:

  • research_agent receives agent.WithExternalTools(...) through graph.WithAgentNodeRunOptions(...).
  • external_tool_gate is a normal graph node. It reads the child agent tool call written by the AgentNode output mapper and calls graph.Interrupt with the toolCallId.
  • The caller supplies the external search result.
  • The graph resumes from the checkpoint with graph.NewResumeCommand().WithResumeMap(...).
  • external_tool_gate stores the caller result as a tool message in graph state.
  • research_agent uses graph.WithAgentNodeInputMapper(...) to project that tool message into graph.StateKeyAgentInputMessage.
  • research_agent runs again with that tool message and writes the final answer.
  • The process keeps one runner session open so later turns reuse the same conversation history.

Run

From the examples/graph module:

export OPENAI_BASE_URL="https://your-openai-compatible-base-url"
export OPENAI_API_KEY="<your api key>"
go run ./agentnode_llmagent_externaltool -model deepseek-v4-flash

Type a user request. When the graph pauses, paste the external tool result at the external_search result> prompt. Type another user request to start the next turn, or type /exit to quit.

The output has this shape:

Session: agentnode-llmagent-externaltool-session-xxx
Type a request to start a turn. Type /exit to quit.
user> Use external search to get user input and say hello+tool result after you receive it

Turn #1: waiting for external_search interrupt.
toolCallId: call_xxx
toolArgs: {"query": "user input"}
checkpointId: checkpoint-xxx
external_search result> 1234

Turn #1: resuming graph.

Final answer:
Hello 1234
user> Use external search to get user input and say hi+tool result after you receive it

Turn #2: waiting for external_search interrupt.
toolCallId: call_xxx
toolArgs: {"query": "user input"}
checkpointId: checkpoint-xxx
external_search result> 56789

Turn #2: resuming graph.

Final answer:
Hi 56789