🧰⛓️‍💥 async-agentic-tools

True asynchronous agentic tools — the model dispatches a tool, gets an immediate acknowledgement, and keeps talking. Results are delivered via callback when they complete — no blocked loops, no dead air.

This is not parallel tool calling (which most agent frameworks already support). Parallel tool calling still blocks the agent loop until every tool in the batch returns. This is true async: the model stays responsive while tools run in the background, and results stream in as they finish.

The demo is built on Strands Agents, but the pattern applies to any agent framework with a tool-calling loop.

Read the blog post for the full explanation of the problem and how this works.

Quick walkthrough video here: "Do async tool calls work now???"

How it works

Three small components layer on top of a standard Strands Agent:

• @tool_async(manager) — Decorator that wraps any tool function. The tool is submitted to a background thread and returns a task ID immediately. Your tool code doesn't change at all.
• AsyncToolManager — Manages a thread pool, tracks pending tasks, and fires a callback when each one completes.
• AsyncAgent — Wraps a Strands Agent to handle result delivery. If the agent is idle when a result arrives, it's delivered immediately. If the agent is busy, results queue up and drain when it finishes.

from strands import Agent
from strands_async_tools import AsyncAgent, AsyncToolManager, tool_async

manager = AsyncToolManager(max_workers=4)

@tool_async(manager)
def slow_research(topic: str) -> str:
    """Research a topic thoroughly."""
    # This runs in a background thread — takes as long as it needs
    time.sleep(15)
    return f"Findings about {topic}..."

agent = Agent(model=model_id, tools=[slow_research])
async_agent = AsyncAgent(agent=agent, manager=manager)
async_agent.send("Research quantum computing")

The framework is about 320 lines of Python across three files in strands_async_tools/.

Prerequisites

• Python 3.14+
• uv (recommended)
• AWS credentials configured for Amazon Bedrock

CLI demo

An interactive chat with three simulated async tools (10-20s delays each) and two synchronous tools from strands-agents-tools (calculator, current_time).

git clone https://github.com/mikegc-aws/async-agentic-tools
cd async-agentic-tools
uv run python demo.py

The default model is Claude Sonnet on Bedrock. Override with the STRANDS_MODEL env var.

Try something like:

You: Research Paris

  [thinking] processing...
I've started researching Paris for you (Task a1b2c3).
I'll let you know as soon as the results come in.

You: What time is it there?

  [thinking] processing...
It's currently 15:32 in Paris (CET, UTC+1).

  [callback] research_topic (a1b2c3) completed in 16482ms — delivering to agent now
  [thinking] processing...
The Paris research just came back! Here are some highlights:
- ...

The async tool dispatches to a background thread and the agent keeps talking. The sync tool (current_time) returns instantly — and the agent knows "there" means Paris. When the research finishes, the result is delivered via callback and the agent speaks it.

Voice mode (experimental)

The voice/ folder contains a voice interface using Amazon Nova Sonic (bidirectional streaming voice model). The agent talks to you through your speakers and listens through your microphone. While you chat, it can delegate complex tasks (web research, file I/O) to a background subagent via @tool_async. Results are injected back into the voice stream — the agent speaks them to you when they're ready.

Voice requirements

• AWS credentials with Bedrock access in us-east-1
• Microphone and speakers
• Model access for amazon.nova-2-sonic-v1:0
• Tavily API key (optional but recommended) — the subagent uses Tavily for web search and extraction. You can get a free API key at app.tavily.com. Without it, the voice agent works but can't do web searches.

Running voice mode

From the repo root:

# With web search (recommended)
TAVILY_API_KEY=tvly-your-key-here uv run python -m voice

# Without web search (still works, just no web research)
uv run python -m voice

Speak into your microphone. The agent responds through your speakers with echo cancellation (LiveKit WebRTC APM). Press Ctrl+C to quit.

Voice env vars

Variable	Default	Description
AWS_REGION	us-east-1	AWS region for Bedrock
NOVA_SONIC_VOICE	tiffany	Nova Sonic voice name
NOVA_SONIC_MODEL	amazon.nova-2-sonic-v1:0	Nova Sonic model ID
SUBAGENT_MODEL	us.anthropic.claude-sonnet-4-20250514-v1:0	Subagent Bedrock model
TAVILY_API_KEY	(none)	Tavily API key for web search
LOG_FILE	voice_debug.log	Debug log path
LOG_LEVEL	WARNING	Console log level

See voice/README.md for more detail on the voice architecture.

Project structure

strands_async_tools/          # The framework (3 files, ~320 lines)
  manager.py                  # AsyncToolManager — thread pool + callbacks
  decorator.py                # @tool_async — decorator for async tools
  agent.py                    # AsyncAgent — callback-driven result delivery

demo.py                       # Interactive CLI demo (simulated async tools)
demo_sync.py                  # Synchronous comparison demo

voice/                        # Experimental voice interface
  voice.py                    # BidiAgent + Nova Sonic + async result injection
  subagent.py                 # Background agent with real tools (web search, file I/O)
  echo_cancel.py              # Echo cancellation (LiveKit WebRTC APM)