dynamic-agent-tools

Dynamic Agent Tools Example

Demonstrates how an AI agent can dynamically extend its tool capabilities at runtime — the core use case for agentic dataflows.

Architecture

Initial topology (1 tool):

Timer (1 Hz) --> Agent --> tool-request --> Echo Tool --> response --> Agent

After adding calculator tool:

                                       ┌── Echo Tool ── response ──┐
Timer (1 Hz) --> Agent -- tool-request─┤                            ├── Agent
                                       └── Calc Tool ── response ──┘

Both tools receive all requests via the shared tool-request topic. Each tool filters for its own "tool" field in the JSON request.

Nodes

agent (agent.py) — Sends JSON tool requests ({"tool": "echo", "message": "..."}) and logs responses. Cycles through echo and calc tasks.

echo-tool (echo_tool.py) — Built-in. Echoes back messages for "tool": "echo".

calc-tool (calc_tool.py) — Dynamically added. Evaluates math expressions for "tool": "calc" requests.

Run

Step 1: Start with echo tool only

pip install dora-rs pyarrow
dora up
dora start examples/dynamic-agent-tools/dataflow.yml --detach --name agent-demo

The agent sends requests. Echo tool responds to echo requests. Calc requests go unanswered (no calc tool yet).

Step 2: Add calculator tool

dora node add --from-yaml examples/dynamic-agent-tools/calc-tool-node.yml --dataflow agent-demo

After wiring the calc tool's response to the agent:

dora node connect --dataflow agent-demo calc-tool/response agent/tool-response

The agent now receives responses from both tools on the same tool-response input (fan-in). Multiple sources can map to the same target input — messages from all sources are interleaved in arrival order.

Step 3: Remove calculator tool

dora node remove agent-demo calc-tool

Back to echo-only. Calc requests go unanswered again.

Step 4: Clean up

dora stop agent-demo
dora down

What This Demonstrates

Feature	How It's Used
Dynamic tool addition	Agent gains new capabilities at runtime
Fan-out to multiple tools	Same tool-request topic, multiple subscribers
Tool-specific filtering	Each tool processes only its own request type
Graceful removal	Removing a tool doesn't affect others
AI agent architecture	Matches the LLM -> Tools service pattern

AI Agent Pattern

This example models the LLM function-calling pattern:

1. Agent decides which tool to call (simulated by cycling through tasks)
2. Agent sends a structured request (JSON with tool field)
3. Tool processes and responds
4. Agent receives and logs the response

In production, the agent node would be an LLM inference node, and tools would be specialized nodes (web search, database query, code execution, etc.) added dynamically based on the conversation context.