pygents 0.3.4

A lightweight async framework for structuring and running AI agents in Python

pygents

A lightweight async framework for structuring and running AI agents in Python. Define tools, queue turns, stream results.

Install

pip install pygents

Requires Python 3.12+.

Example

import asyncio
from pygents import Agent, Turn, tool

@tool()
async def greet(name: str) -> str:
    return f"Hello, {name}!"

async def main():
    agent = Agent("greeter", "Greets people", [greet])
    # Use kwargs:
    await agent.put(Turn("greet", kwargs={"name": "World"}))
    # Or positional args:
    await agent.put(Turn("greet", args=["World"]))

    async for turn, value in agent.run():
        print(value)  # "Hello, World!"

asyncio.run(main())

Tools are async functions. Turns say which tool to run and with what args. Agents process a queue of turns and stream results. The loop exits when the queue is empty.

Features

• Streaming — agents yield (turn, value) as results are produced
• Inter-agent messaging — agents can send turns to each other
• Dynamic arguments — callable positional args and kwargs evaluated at runtime
• Timeouts — per-turn, default 60s
• Per-tool locking — opt-in serialization for shared state (lock is acquired inside the tool wrapper, so turn-level hooks run outside the tool lock)
• Fixed kwargs — decorator kwargs (e.g. @tool(permission="admin")) are merged into every invocation; call-time kwargs override
• Hooks — @hook(hook_type, lock=..., **fixed_kwargs) decorator; hooks stored as a list and selected by type; turn, agent, tool, and memory hooks; same fixed_kwargs and lock options as tools
• Serialization — to_dict() / from_dict() for turns and agents

Design decisions

• Turn identity: Turn instances no longer have a built-in uuid. If you need identifiers, store them yourself in metadata or wrap Turn in a higher-level domain object.

• Turn arguments: Turn.__init__ takes args before kwargs, and metadata is the final parameter:

  Turn(
      "tool_name",
      args=[...],
      kwargs={...},
      timeout=...,
      metadata={...},
      hooks=[...],
  )
  

  This keeps positional arguments explicit while reserving metadata purely for user-level annotations. start_time, end_time, stop_reason, and output are set by the framework during execution—they are not constructor parameters.

• Agent serialization and current turn: Agent.to_dict() includes a current_turn key when the agent is in the middle of run() (the turn being executed). That turn is already off the queue, so without it a snapshot would lose the in-flight work. from_dict() restores it; the next run() consumes the restored current turn first, then the queue. So you can save agent state at any time and get a faithful snapshot (config, queue, and current turn if any).

• Tool and hook metadata timing: ToolMetadata and HookMetadata include start_time and end_time (datetime | None). They are set on the metadata instance when the tool or hook runs (start at entry, end in a finally). So the same metadata object is updated each run; dict() includes ISO strings for serialization.

• Hook protocol: Registered hooks conform to a Hook protocol (like Tool), with metadata (HookMetadata: name, description, and run timing), hook_type, fn, and lock. The decorator sets metadata (from __name__ and __doc__), fn, and lock; HookRegistry.register() sets hook_type. Raw async callables are typed as Callable[..., Awaitable[None]].

• Hook decorator: @hook(hook_type, lock=False, **fixed_kwargs) mirrors the tool decorator: keyword arguments are merged into every invocation (call-time overrides), and lock=True uses an asyncio lock to serialize hook invocations. Pass a list of types (e.g. @hook([TurnHook.BEFORE_RUN, AgentHook.AFTER_TURN])) to reuse one hook for several events; multi-type hooks must accept *args, **kwargs since different types receive different arguments.

• Tool lock in tool layer: The tool lock is acquired inside the @tool wrapper, not in the turn. So the lock covers only the tool’s own execution (including its BEFORE_INVOKE / ON_YIELD / AFTER_INVOKE hooks). Turn-level hooks (BEFORE_RUN, AFTER_RUN, ON_TIMEOUT, ON_ERROR) run outside the tool lock; hooks that need serialization use their own lock=True.

• Memory hooks: Memory has no compact callback. It supports MemoryHook.BEFORE_APPEND and MemoryHook.AFTER_APPEND only. Hooks are stored as list[Hook] (like Agent/Turn), filtered by type when running. BEFORE_APPEND and AFTER_APPEND hooks receive (items,) — the current items as a list (read-only). Serialization uses the same by-type-by-name shape as Agent/Turn; from_dict() resolves names from HookRegistry.

Docs

Full documentation: uv run mkdocs serve. MkDocs is an optional dependency—install with pip install -e ".[docs]" (or use uv run as above) so the library itself does not depend on it.