autourgos-agent

A self-contained, general-purpose LLM agent for the Autourgos framework — reasoning and acting in one clean loop, with any OpenAI-compatible LLM you already have.

The agent alternates between Thought (reasoning about what to do next) and Action (calling a tool), looping until it has enough information to give a Final Answer.

Fully self-contained — zero third-party runtime dependencies beyond Python 3.10+ (only autourgos-core, itself a zero-dependency stdlib utility library shared across the framework). No forced LLM SDK, no forced vector store, no hidden network calls at import time. Bring your own LLM wrapper.

Area	What you get
Core loop	Thought → Action → Observation, JSON-driven; native structured tool-calling mode too; parallel tool calls (thread pool / asyncio.gather); async everywhere — every method has an a-prefixed twin
Built in, zero extra install	Lazy-loaded toolboxes for large tool catalogs; per-iteration file/callback injection (screenshots, live data); run history to disk with secrets auto-redacted; pause/resume an in-flight run from any thread; auto-summarizing scratchpad, retry-with-backoff, timeouts
Extensible	CallbackHandler middleware with 11 lifecycle hooks; sync or async hooks, mixed freely, from either loop; approval callbacks for human-in-the-loop / safety gates; works with autourgos-hcix, autourgos-skills, and any hand-written middleware
Any LLM	OpenAI, Groq, Together AI, Mistral, DeepSeek, Perplexity; Ollama / LM Studio / vLLM — fully local, no API key; anything with .invoke() / .ainvoke() — no lock-in

---

Why use this?

Almost every major LLM provider today exposes an OpenAI-compatible API. autourgos-agent was designed with this in mind — it works with any LLM that has .invoke() and .ainvoke() methods. One agent, any provider:

Provider	Notes
OpenAI	gpt-4o, gpt-4o-mini, ...
Groq	Llama 3, Mixtral, Gemma
Together AI	100+ open-source models
Mistral AI	mistral-large, codestral
DeepSeek	deepseek-chat, deepseek-reasoner
Perplexity	sonar — web-connected
Ollama	local models, no internet
LM Studio	local models, GUI-based
vLLM	self-hosted, high throughput

You are not locked to a single provider.

What does it do?

The agent receives a task and a list of tools. It then iterates:

1. Think — what information do I need? which tool should I call?
2. Act — call the tool, get the result
3. Observe — add the result to the scratchpad, repeat

This continues until the agent has a final answer or hits the iteration/time limit.

---

Table of Contents

Getting started

• Install
• Quick Start
• How the Agent Loop Works
• Defining Tools
• Works With Any LLM

Running the agent

• Async Agent
• Parallel Tool Calls
• Native Tool Calling
• Verbose Mode
• Memory
• Approval Callback

Extending the agent

• Middleware / Callbacks
  • Middleware Integration Contract
• on_agent_start Shortcut
• Toolboxes (Lazy-Loaded Tool Groups)
• Pre-Iteration Files & Callbacks
• Run History
• Pause & Resume

Operating it in production

• Testing
• Context Manager
• Time and Iteration Limits
• Scratchpad Size Limits
  • Auto-Summarizing Scratchpad
• LLM Call Retries
• Custom System Prompt

Reference

• Constructor Reference
• Tool Dict Reference
• What the Agent Returns
• Exceptions
• v1 Backward Compatibility

---

Install

pip install autourgos-agent

No required third-party runtime dependencies (only autourgos-core, zero-dep itself). Bring your own LLM wrapper:

pip install autourgos-openaichat   # Chat Completions API
# or
pip install autourgos-responses    # OpenAI Responses API

Requires Python 3.10+.

---

Quick Start

from autourgos_agent import Agent, tool
from autourgos_openaichat  import OpenAIChatModel

# 1. Define a tool
@tool
def calculator(a: float, b: float) -> float:
    """Add two numbers together."""
    return a + b

# 2. Create the agent
agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    verbose=True,
)
agent.add_tools(calculator)

# 3. Run
result = agent.invoke("What is 123 + 456?")
print(result)
# 579

(Prefer writing tools by hand instead? A plain dict — {"name", "description", "parameters", "func"} — still works exactly the same way; see Defining Tools below.)

Expected verbose output (LangChain-flavored Thought/Action/Observation trace):

> Starting Agent...

Thought: I need to add 123 and 456. I'll use the calculator tool.
Action: calculator
Action Input: {'a': 123, 'b': 456}
Observation: 579.0
Thought: I have the result from the calculator.
Final Answer: 123 + 456 = 579

> Agent finished.

(Shown here without colour; in a real terminal > Starting Agent..., > Agent finished., and Final Answer: print in bold green, Action: / Action Input: in yellow, Observation: in blue/cyan, Thought: in cyan, and Parse Error: in red.)

---

How the Agent Loop Works

Each iteration the agent produces a JSON object:

{
  "thought": "I need to search for the latest Python version.",
  "actions": [
    {"action": "search", "action_input": {"query": "latest Python version 2025"}}
  ],
  "final_answer": null
}

Rules the LLM must follow (enforced by the prompt):

• If tools are needed → fill actions, set final_answer to null
• If the answer is ready → fill final_answer, set actions to []
• Never set both actions and final_answer at the same time
• Multiple tools can be called in one step if they are independent

The agent collects tool results into a scratchpad that is passed back to the LLM at each step so it always has full context of what was tried.

---

Defining Tools

There are two ways to define a tool. Both produce the same shape under the hood and can be freely mixed on the same agent.

Recommended: the @tool decorator

Decorate a type-hinted function and the name, description, and JSON-Schema parameters are inferred automatically — no dict to write by hand:

from autourgos_agent import tool

@tool
def get_weather(city: str, unit: str = "celsius") -> str:
    """Get the current weather for a city.

    Args:
        city: City name, e.g. Tokyo
        unit: celsius or fahrenheit
    """
    # Replace with real API call
    return f"The weather in {city} is 22°{unit[0].upper()} and sunny."

agent.add_tools(get_weather)

• name defaults to the function's name (override with @tool(name=...)).
• description defaults to the first line of the docstring (override with @tool(description=...)).
• parameters are inferred from type hints (str/int/float/bool/list/dict → the matching JSON-Schema type; parameters without a default are marked required). Per-parameter descriptions are parsed from a Google-style Args: section if present. Override entirely with @tool(parameters={...}) if you need something the inference can't express.
• The decorated function stays directly callable — get_weather("Tokyo") still works outside the agent, e.g. in your own tests.

# Overriding the inferred name/description:
@tool(name="calculator", description="Add two numbers together.")
def add(a: float, b: float) -> float:
    return a + b

Alternative: a plain dict

For full manual control (or if you're integrating an existing tool spec), a tool can still be a plain Python dict with these keys:

Key	Type	Required	Description
name	str	yes	Tool name (used by the LLM to call it)
description	str	yes	What the tool does — shown to the LLM
parameters	dict	recommended	JSON-Schema object describing the inputs
func	callable	yes	Python function to call

weather_tool = {
    "name": "get_weather",
    "description": "Get the current weather for a city.",
    "parameters": {
        "type": "object",
        "properties": {
            "city": {"type": "string",  "description": "City name, e.g. Tokyo"},
            "unit": {"type": "string",  "description": "celsius or fahrenheit"},
        },
        "required": ["city"],
    },
    "func": get_weather,
}

Duck-typed tool objects

A tool doesn't have to be a dict at all — any object exposing .name, .description, .parameters, and .func (or .function) as attributes instead of dict keys works the same way in both tool_calling_mode="prompt" and "native":

class WeatherTool:
    name = "get_weather"
    description = "Get the current weather for a city."
    parameters = {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
    func = staticmethod(get_weather)

agent.add_tools(WeatherTool())

Useful when wrapping an existing tool class from another library instead of reshaping it into a dict.

Adding tools

@tool-decorated functions and plain dicts both work the same way with add_tools / the constructor, and can be mixed freely:

# One at a time
agent.add_tools(get_weather)

# Multiple at once (mix @tool and dict tools freely)
agent.add_tools(get_weather, calculator_tool, search_tool)

# From a list
agent.add_tools([get_weather, calculator_tool])

# Via constructor
agent = Agent(llm=llm, tools=[get_weather, calculator_tool])

---

Works With Any LLM

Change the llm= argument to switch providers. Everything else stays the same.

OpenAI

from autourgos_openaichat import OpenAIChatModel

agent = Agent(llm=OpenAIChatModel(model="gpt-4o", api_key="sk-..."))

Groq (very fast, free tier)

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="llama3-70b-8192",
        api_key="gsk_...",
        base_url="https://api.groq.com/openai/v1",
    )
)

Ollama (fully local, no internet, no API key)

ollama pull llama3

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="llama3",
        api_key="ollama",
        base_url="http://localhost:11434/v1",
    )
)

Together AI

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="meta-llama/Llama-3-70b-chat-hf",
        api_key="...",
        base_url="https://api.together.xyz/v1",
    )
)

Mistral AI

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="mistral-large-latest",
        api_key="...",
        base_url="https://api.mistral.ai/v1",
    )
)

DeepSeek

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="deepseek-chat",
        api_key="...",
        base_url="https://api.deepseek.com/v1",
    )
)

OpenAI Responses API (autourgos-responses)

from autourgos_responses import OpenAIResponse

agent = Agent(llm=OpenAIResponse(model="gpt-4o"))

LM Studio (local GUI)

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="local-model",
        api_key="lm-studio",
        base_url="http://localhost:1234/v1",
    )
)

vLLM (self-hosted)

from autourgos_openaichat import OpenAIChatModel

agent = Agent(
    llm=OpenAIChatModel(
        model="meta-llama/Meta-Llama-3-8B-Instruct",
        api_key="EMPTY",
        base_url="http://your-server:8000/v1",
    )
)

---

Async Agent

All agent methods have an async counterpart.

import asyncio
from autourgos_agent import Agent
from autourgos_openaichat  import OpenAIChatModel

agent = Agent(llm=OpenAIChatModel(model="gpt-4o"))
agent.add_tools(weather_tool, calculator_tool)

async def main():
    result = await agent.ainvoke("What is the weather in Tokyo and what is 99 * 3?")
    print(result)
    # The weather in Tokyo is 22°C and sunny. 99 × 3 = 297.

asyncio.run(main())

Async tools (coroutine functions) are also supported:

import httpx

@tool
async def search(query: str) -> str:
    """Search the web."""
    async with httpx.AsyncClient() as client:
        r = await client.get(f"https://api.example.com/search?q={query}")
        return r.text
# async function — works with ainvoke(), same as a plain-dict "func" would.

---

Parallel Tool Calls

The LLM can call multiple tools in a single step when they don't depend on each other. The agent runs the approved ones concurrently (a ThreadPoolExecutor for invoke(), asyncio.gather for ainvoke()) and collects all results before the next LLM call. invoke()'s thread pool is capped at Agent.MAX_TOOL_WORKERS (default 8) regardless of how many tool calls the model requests in one step, so a step with more approved calls than that queues the rest rather than spawning one thread per call.

agent = Agent(llm=OpenAIChatModel(model="gpt-4o"))
agent.add_tools(weather_tool, calculator_tool, search_tool)

result = agent.invoke(
    "What is the weather in Paris and London, and what is 250 * 4?"
)
print(result)
# The weather in Paris is 18°C cloudy, London is 15°C rainy. 250 × 4 = 1000.

The LLM produces three tool calls in one step:

{
  "thought": "I can fetch both cities' weather and compute the multiplication in parallel.",
  "actions": [
    {"action": "get_weather", "action_input": {"city": "Paris"}},
    {"action": "get_weather", "action_input": {"city": "London"}},
    {"action": "calculator",  "action_input": {"a": 250, "b": 4}}
  ],
  "final_answer": null
}

For native structured tool-calling instead of JSON-in-text parsing, see Native Tool Calling below.

---

Native Tool Calling

tool_calling_mode="native" swaps the JSON-in-text agent loop for the LLM's own structured tool-calling — invoke_with_tools()/ainvoke_with_tools() on OpenAIChatModel/OpenAIResponse. No regex JSON parsing — tool calls come back as structured data straight from the API. Multiple tool calls in one turn run concurrently here too, the same way "prompt" mode does.

from autourgos_agent import Agent, tool
from autourgos_openaichat import OpenAIChatModel

@tool
def weather_tool(city: str) -> str:
    """Get the current weather for a city."""
    return f"{city}: 18°C, cloudy"

@tool
def calculator_tool(a: float, b: float) -> float:
    """Multiply two numbers."""
    return a * b

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    tool_calling_mode="native",
)
agent.add_tools(weather_tool, calculator_tool)

result = agent.invoke("What is the weather in Paris, and what is 250 * 4?")
print(result)

Requires the LLM to implement invoke_with_tools()/ainvoke_with_tools() — both OpenAIChatModel and OpenAIResponse do. Passing tool_calling_mode="native" with an LLM that doesn't (a plain duck-typed .invoke()-only object, or the default BaseLLM.invoke_with_tools() stub) raises a RuntimeError immediately, naming the LLM class and what's missing — it never silently falls back to prompt mode.

Trade-offs versus "prompt" mode (the default):

• More reliable: no AgentParseError from malformed JSON, since the API returns structured tool calls directly.
• No visible "Thought" per step: when the model also calls tools, the wrapper doesn't currently return accompanying reasoning text alongside the tool calls, so on_iteration/logger.thought() aren't fired on tool-call turns — only on the final answer.
• Conversation state is a real multi-turn message list, not the single rendered scratchpad string "prompt" mode uses. agent.scratchpad is still kept up to date as a human-readable trace (for middleware that reads it), but it isn't what's actually sent to the LLM in this mode.

---

Verbose Mode

Enable verbose=True to print every step to stdout.

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    verbose=True,
)
agent.add_tools(weather_tool)
result = agent.invoke("What is the weather in Sydney?")

Output (LangChain-flavored Thought/Action/Observation trace):

> Starting Agent...

Thought: I need to get the weather in Sydney using the get_weather tool.
Action: get_weather
Action Input: {'city': 'Sydney'}
Observation: The weather in Sydney is 25°C and sunny.
Thought: I have the weather information for Sydney.
Final Answer: The weather in Sydney is 25°C and sunny.

> Agent finished.

Enable full_output=True to also print the raw LLM JSON at each step — useful for debugging prompt or parse issues:

agent = Agent(llm=llm, verbose=True, full_output=True)

---

Memory

Attach a memory backend to persist conversation history across calls.

from autourgos_agent import Agent, MemoryProtocol
from typing import Dict, List

class SimpleMemory(MemoryProtocol):
    def __init__(self):
        self._history: List[Dict[str, str]] = []

    def add_user_message(self, message: str) -> None:
        self._history.append({"role": "user", "content": message})

    def add_assistant_message(self, message: str) -> None:
        self._history.append({"role": "assistant", "content": message})

    def get_history(self) -> List[Dict[str, str]]:
        return list(self._history)

memory = SimpleMemory()
agent  = Agent(llm=llm, memory=memory)
agent.add_tools(search_tool)

result1 = agent.invoke("Search for the capital of France.")
print(result1)
# The capital of France is Paris.

result2 = agent.invoke("What city did I just ask about?")
print(result2)
# You asked about Paris, the capital of France.

---

Approval Callback

Require human (or programmatic) approval before any tool is executed.

def require_approval(tool_name: str, tool_input: dict) -> bool:
    print(f"\n[Approval required] Tool: {tool_name}")
    print(f"Input: {tool_input}")
    answer = input("Allow? (y/n): ").strip().lower()
    return answer == "y"

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    approval_callback=require_approval,
)
agent.add_tools(delete_file_tool)
result = agent.invoke("Delete the temp folder.")

If the callback returns a falsy value, the tool is skipped and the agent sees:

Observation: Tool call was denied by the approval callback.

A denial still fires both on_tool_start and on_tool_end (with that same "denied by the approval callback" message as the result) — middleware that tracks tool calls by pairing start/end events sees a consistent pair either way, never a start with no matching end.

Use this to implement human-in-the-loop, audit logging, or safety checks for destructive tools.

approval_callback can also be an async def — ainvoke() awaits it if it returns an awaitable (e.g. to wait on a Slack approval), and a plain sync callback still works unchanged in both invoke() and ainvoke():

async def require_approval(tool_name: str, tool_input: dict) -> bool:
    return await ask_on_slack(tool_name, tool_input)

agent = Agent(llm=OpenAIChatModel(model="gpt-4o"), approval_callback=require_approval)
agent.add_tools(delete_file_tool)
result = await agent.ainvoke("Delete the temp folder.")

Note: invoke() (the sync entrypoint) only supports a sync approval_callback. Passing an async def callback to invoke() raises a TypeError right away (naming the problem and pointing at the fix) instead of silently approving every tool — call agent.ainvoke() instead, or use a plain sync callback.

---

Middleware / Callbacks

Register event hooks to observe the agent without modifying it. CallbackHandler exposes 11 hooks in total: on_agent_start, on_agent_end, on_agent_error, on_tool_start, on_tool_end, on_tool_error, on_iteration_start, on_before_iteration, on_iteration, on_llm_end, and on_parse_error. Every hook may receive an agent=<Agent instance> kwarg (older handlers that don't accept it still work).

on_agent_error also fires on cancellation. invoke()/ainvoke() catch BaseException, not just Exception, so on_agent_error is guaranteed to fire — exactly once, with a bare re-raise afterward — for an ordinary exception, a cancelled async run (asyncio.CancelledError), or a KeyboardInterrupt/SystemExit mid-run. Do cleanup that must always happen (removing tools/prompt blocks your middleware injected, stopping a listener, flushing a log, deleting temp files) in on_agent_error, not just on_agent_end — otherwise a cancelled run silently skips it. If your handler treats on_agent_error as "a real application failure occurred" (e.g. alerting), check isinstance(error, (asyncio.CancelledError, KeyboardInterrupt, SystemExit)) to distinguish cancellation from an actual error.

from autourgos_agent import Agent, CallbackHandler

class MyLogger(CallbackHandler):

    def on_agent_start(self, query: str, **kwargs) -> None:
        print(f"Agent started with query: {query}")

    def on_agent_end(self, result: str, **kwargs) -> None:
        print(f"Agent finished: {result}")

    def on_tool_start(self, tool_name: str, tool_input: dict, **kwargs) -> None:
        print(f"Calling tool: {tool_name} with {tool_input}")

    def on_tool_end(self, tool_name: str, result: str, **kwargs) -> None:
        print(f"Tool {tool_name} returned: {result[:100]}")

    def on_iteration(self, iteration: int, thought: str, **kwargs) -> None:
        print(f"Iteration {iteration} — thought: {thought}")

    def on_parse_error(self, iteration: int, raw_response: str, **kwargs) -> None:
        print(f"Parse error at step {iteration}: {raw_response[:100]}")

    # Extra hooks (v1.1.0+) — all optional, all receive `agent=` too:
    def on_agent_error(self, error: Exception, **kwargs) -> None:
        print(f"Agent error: {error}")

    def on_tool_error(self, tool_name: str, error: Exception, **kwargs) -> None:
        print(f"Tool {tool_name} raised: {error}")

    def on_iteration_start(self, iteration: int, **kwargs) -> None:
        print(f"Starting iteration {iteration}")

    def on_llm_end(self, response, **kwargs) -> None:
        print(f"LLM responded: {str(response)[:100]}")
        # kwargs also carries raw=<untouched raw LLM response> plus, when the
        # wrapper exposes it: provider_used, input_tokens, output_tokens,
        # total_cost, latency_ms (autourgos-openaichat/-responses' dict shape)
        # or total_tokens (from a native SDK response's .usage) — useful for
        # cost/usage-tracking middleware without reaching into agent.llm.
        if "total_cost" in kwargs:
            print(f"  cost so far: ${kwargs['total_cost']}")


agent = Agent(llm=llm, middleware=[MyLogger()])
agent.add_tools(weather_tool)
result = agent.invoke("Weather in Berlin?")

You can also add middleware after construction:

agent.add_middleware(MyLogger())

Narrating middleware activity in the verbose trace

By default, verbose=True only shows the core loop (Thought/Action/Observation) -- middleware changing the agent's tools, scratchpad, or prompt behind the scenes is otherwise invisible. Middleware can narrate what it's doing into the same trace via agent.logger.middleware(source, message):

class MyLogger(CallbackHandler):
    def on_agent_start(self, query: str, agent=None, **kwargs) -> None:
        logger = getattr(agent, "logger", None)
        if logger:
            logger.middleware("MyLogger", "Doing something worth narrating.")

Printed in magenta with a [Source] prefix so it's unambiguous which middleware produced the line, e.g.:

[Toolbox] Exposed toolbox 'search_tools' to agent.
[Summarizer] Compressed scratchpad (iteration 5, was 15,320 chars).

Use getattr(agent, "logger", None) (not a direct import of AgentLogger) so your middleware doesn't crash if it's ever attached to something other than a Agent, and does nothing when verbose=False. The built-in summarizer, toolbox, and pre-iteration runtimes (see Auto-Summarizing Scratchpad, Toolboxes, and Pre-Iteration Files & Callbacks — all three implemented inline, not as middleware, but narrate the same way) and sibling middleware packages (e.g. autourgos-hcix, autourgos-skills) all use this same pattern to narrate their own actions.

Middleware Integration Contract

These are the three pieces of surface area sibling middleware (autourgos-hcix, autourgos-skills, and anything else you write) can rely on. This is the official, stable contract — treat it as public API.

agent.scratchpad (str) A real, live instance attribute, not just a local loop variable. It is updated in place on every iteration of invoke()/ainvoke(), so a callback handler (or any other code holding a reference to the agent) can read it while the loop is still running, not only after it finishes. It's reset to "" at the start of every invoke()/ainvoke() call, so calling invoke() twice on the same agent instance never leaks the previous run's scratchpad into the new one.

class ScratchpadWatcher(CallbackHandler):
    def on_iteration_start(self, iteration, agent=None, **kwargs):
        print(f"[iter {iteration}] scratchpad so far:\n{agent.scratchpad}")

agent.current_query (str) Set once, at the start of every invoke()/ainvoke() call, to the query being worked on. Lets middleware answer "what is this agent doing right now?" without threading the query through every hook signature.

on_tool_start / on_tool_end are always paired Every on_tool_start for a given tool call is followed by exactly one on_tool_end for that same call — including when approval_callback denies it (see Approval Callback). Middleware that tracks in-flight tool calls (e.g. an "active calls" gauge, or a tracing span opened on start and closed on end) can rely on this pairing without special-casing denials.

on_before_iteration(iteration, agent=None, **kwargs) Called once per loop iteration, right before the LLM is invoked for that iteration. If a handler returns a dict, its keys are merged into the self.llm.invoke() / self.llm.ainvoke() call for that iteration only — it is not persisted to later iterations. If multiple handlers return dicts, later handlers win on key conflicts. Returning None (the default no-op, same as every other hook) changes nothing.

class TemperatureOverride(CallbackHandler):
    def on_before_iteration(self, iteration, agent=None, **kwargs):
        # only lower the temperature on the first iteration
        if iteration == 1:
            return {"temperature": 0.0}
        return None

This is how middleware can, for example, inject a trace id, override a sampling parameter, or attach per-call metadata without the agent needing to know anything about the specific middleware doing it.

Hooks may be sync or async, from either invoke() or ainvoke() Define a hook as a plain def or an async def — both work from both loops:

• From ainvoke() (the async loop), a sync hook runs on a background thread rather than the event-loop thread, so a blocking call inside it (an LLM request, a file write, time.sleep, anything) doesn't stall the loop for every other concurrent ainvoke() call sharing that thread. An async hook is awaited directly.
• From invoke() (the sync loop), a sync hook is called directly, same as always; an async hook is driven to completion with its own short-lived event loop.

class RemoteAudit(CallbackHandler):
    async def on_iteration_start(self, iteration, agent=None, **kwargs):
        await audit_client.log(agent.current_query, iteration)

You don't need to pick one style for a whole handler — different hooks on the same class can mix sync and async freely.

---

on_agent_start Shortcut

For the common case of wanting one function to run every time an agent starts, without writing a full CallbackHandler subclass:

def log_start(query: str) -> None:
    print(f"Starting: {query}")

agent = Agent(llm=llm, on_agent_start=log_start)

on_agent_start may be a plain function or an async def — both work from invoke() and ainvoke(). It's called as fn(query), or fn(query, agent=self) if the function's signature accepts an agent kwarg:

async def log_start(query: str, agent=None) -> None:
    await audit_client.log(query, agent.current_query)

agent = Agent(llm=llm, on_agent_start=log_start)

This is internally just sugar for agent.add_middleware(...) wrapping your function in a CallbackHandler — it fires alongside (after) any handlers passed via middleware=.

---

Toolboxes (Lazy-Loaded Tool Groups)

toolbox= groups tools under a name + description that stay hidden from the agent's prompt until it actually needs them — only each toolbox's name and one-line description are shown upfront. The agent calls the built-in expose_toolbox(name) (or expose_tool(tool_name) for a single tool) meta-tool at runtime to load the real tools. Useful when you have many tools across several domains but a given run only needs a handful — keeps the context window clean instead of dumping every tool's full schema into the prompt every time.

from autourgos_core import tool, Toolbox
from autourgos_agent import Agent
from autourgos_openaichat import OpenAIChatModel

@tool
def web_search(query: str) -> str:
    """Search the web."""
    return f"Results for: {query}"

@tool
def scrape_url(url: str) -> str:
    """Scrape a web page."""
    return f"Scraped content of {url}"

@tool
def run_query(sql: str) -> str:
    """Run a SQL query against the database."""
    return f"Rows for: {sql}"

@tool
def list_tables() -> str:
    """List all tables in the database."""
    return "users, orders, products"

web = Toolbox(name="web", description="Web search and page scraping tools.",
              tools=[web_search, scrape_url])
db  = Toolbox(name="database", description="SQL query tools.",
              tools=[run_query, list_tables])

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    toolbox=[web, db],
)

result = agent.invoke("Find the latest Python release and save it to the DB")

The agent first sees only:

## Dynamic Toolboxes
- **web**: Web search and page scraping tools.
- **database**: SQL query tools.

Then, when it decides it needs one, it calls expose_toolbox("web") (or expose_tool("web_search") for just that one tool), which registers the real tool(s) on the agent and injects their full schemas into the prompt for the rest of that run. Everything toolbox-related — meta-tools, injected schemas, displaced/restored tools — is automatically cleaned up at on_agent_end/on_agent_error, so the agent returns to its original state before the next invoke().

You can also mix tools= (always visible) with toolbox= (lazy-loaded) on the same agent:

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    tools=[calculator],       # always in the prompt
    toolbox=[web, db],        # hidden until expose_toolbox()/expose_tool() is called
)

toolbox= is built directly into the agent loop, not middleware. Register a toolbox dynamically after construction with agent.add_toolbox(...):

agent = Agent(llm=OpenAIChatModel(model="gpt-4o"))
agent.add_toolbox("web", "Web search and page scraping tools.", [web_search, scrape_url])

---

Pre-Iteration Files & Callbacks

Run a callback and/or inject files (e.g. a fresh screenshot) before every agent iteration — useful for computer-use / vision agents, live data feeds, or anything that needs to refresh before each step.

Built directly into the agent loop via pre_iteration_callback=/ pre_iteration_files= on Agent() — same pattern as history=/ summarize_every=. Not middleware: no CallbackHandler to write, no middleware=[...] to wire up, zero overhead when unset.

from autourgos_agent import Agent
from autourgos_openaichat import OpenAIChatModel

SCREENSHOT = "/tmp/screen.png"

def capture(iteration: int) -> None:
    take_screenshot(SCREENSHOT)  # your own screenshot function

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    pre_iteration_callback=capture,
    pre_iteration_files=SCREENSHOT,
    image_quality="low",   # downscale to <=512px, JPEG q60 -- ~85 tokens flat
)
result = agent.invoke("Click the 'Submit' button on screen.")

pre_iteration_files= also accepts a callable that returns a path (or list of paths) for dynamic/per-iteration file names, and non-image files are passed through unmodified:

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    pre_iteration_files=lambda iteration: f"/tmp/screen_{iteration}.png",
)

image_quality options: "auto" (default, no change), "high" (no resize, detail="high"), "medium" (≤768px, JPEG q70), "low" (≤512px, JPEG q60), or an int 1–100 (JPEG quality directly). Resizing requires Pillow: pip install 'autourgos-agent[images]' — without it, only the detail= hint is applied (still saves tokens on the OpenAI side) and the original image is sent unresized.

The callback can be sync or async, and runs safely from both invoke() (directly) and ainvoke() (offloaded to a worker thread so it never blocks the event loop). A callback that raises is logged (at ERROR) and does not stop the agent run.

Run multiple callbacks with SEQUENTIAL (one after another) or PARALLEL (concurrently — sync callbacks in a thread pool, async ones as asyncio tasks):

from autourgos_agent import SEQUENTIAL, PARALLEL

def log_step(iteration: int) -> None:
    print(f"Iteration {iteration} starting")

async def refresh_cache(iteration: int) -> None:
    await cache.refresh()

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    pre_iteration_callback=SEQUENTIAL[capture, log_step],      # capture, then log, in order
    # or: pre_iteration_callback=PARALLEL[capture, refresh_cache],  # both at once
    pre_iteration_files=SCREENSHOT,
)

---

Run History

history= records every run to a Markdown + JSON file pair on disk — thoughts, tool calls, observations, and the final answer — written directly by the agent loop (not middleware), with secret-shaped values (API keys, bearer tokens, JWTs, ...) automatically redacted before writing.

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    history="./agent_runs",   # folder is created if it doesn't exist
)
agent.add_tools(search_tool)

result = agent.invoke("Research the latest AI news.")
# writes ./agent_runs/Task_<timestamp>_<uid>.md and .json

The Markdown file is a human-readable transcript of the run (thought/action/ observation trace plus the final answer); the JSON file is the same data in a structured form for programmatic use. history=None (the default) disables recording entirely — no files are written.

---

Pause & Resume

agent.pause(reason=None) / agent.resume() / agent.is_paused give you an in-process, thread-safe way to pause a running agent and later hand control back to it — the run blocks at its next iteration boundary (before the next LLM call, never mid-tool-call or mid-LLM-call) until resume() is called. Works from both invoke() and ainvoke(), and in both tool_calling_mode="prompt" and "native".

agent = Agent(llm=my_llm)

# From another thread (or another asyncio task):
agent.pause(reason="waiting for human review")
...
agent.resume()

• Call pause()/resume() from any thread — they don't have to be called from the thread running invoke()/ainvoke().
• Calling pause() before invoke()/ainvoke() starts means that run begins already paused — it blocks before its first iteration. This is a valid way to start an agent pre-paused, not a bug.
• resume() without a prior pause() is a no-op.
• max_execution_time excludes time spent paused — pausing an agent (e.g. to wait for a human) never counts against its execution-time budget.
• A middleware hook can pause the agent it's attached to just as easily as external code:

class PauseForApproval(CallbackHandler):
    def on_iteration_start(self, iteration, agent=None, **kwargs):
        if needs_human_review(agent):
            agent.pause(reason="needs review")

Two new CallbackHandler hooks narrate pause/resume to middleware:

class PauseLogger(CallbackHandler):
    def on_agent_pause(self, iteration, reason, agent=None, **kwargs):
        print(f"Paused at iteration {iteration}: {reason}")

    def on_agent_resume(self, iteration, paused_duration, agent=None, **kwargs):
        print(f"Resumed after {paused_duration:.1f}s")

Out of scope for this feature (deliberately): resuming a paused run in a different process or after the original one has exited — this is an in-process mechanism, not a serialized/checkpointed one. Pausing mid-tool-call is also not supported — a pause only ever takes effect at the next iteration boundary.

---

Testing

autourgos_agent.testing ships make_test_agent() — a shared test fixture that builds a real, fully-functional Agent wired to a scripted fake LLM, with zero network calls. Use it in your own tests instead of hand-rolling a fake agent (a hand-rolled fake's shape can silently drift from the real Agent and hide real bugs):

import json
from autourgos_agent.testing import make_test_agent

agent = make_test_agent(responses=[
    json.dumps({"thought": "thinking", "actions": [], "final_answer": "42"}),
])
result = agent.invoke("what is the answer?")
assert result == "42"
assert agent.llm.call_count == 1

make_test_agent() accepts responses (a list of raw JSON-text canned LLM replies in the {thought, actions, final_answer} format), and optional tools, memory, middleware, max_iterations, and any other Agent constructor kwarg. If tools is omitted, a harmless echo tool is attached automatically so agent.invoke() works out of the box.

---

Context Manager

The agent implements both sync and async context managers. They automatically close the LLM's HTTP client when the block exits.

with Agent(llm=OpenAIChatModel(model="gpt-4o")) as agent:
    agent.add_tools(calculator_tool)
    result = agent.invoke("What is 7 * 8?")
    print(result)
    # 56
# LLM client closed here

Async:

import asyncio
from autourgos_openaichat import OpenAIChatModel

async def main():
    async with Agent(llm=OpenAIChatModel(model="gpt-4o")) as agent:
        agent.add_tools(calculator_tool)
        result = await agent.ainvoke("What is 12 ** 2?")
        print(result)
        # 144

asyncio.run(main())

---

Time and Iteration Limits

Prevent runaway agents with hard limits.

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    max_iterations=10,       # stop after 10 Thought → Action → Observe cycles
    max_execution_time=30.0, # stop after 30 seconds wall-clock time
)
agent.add_tools(search_tool)

try:
    result = agent.invoke("Research the entire history of the internet.")
except AgentTimeoutError:
    ...  # 30s wall-clock elapsed without a final answer
except AgentMaxIterationsError:
    ...  # 10 iterations elapsed without a final answer

You can also override max_iterations per call:

result = agent.invoke("Quick question: capital of Japan?", max_iterations=3)

max_execution_time is rechecked immediately after every blocking LLM call, tool wait, and approval-callback call returns (not just once per iteration), and an in-flight async LLM call is actually cancelled at its next await point via asyncio.wait_for. It still can't force-stop a hanging synchronous call already in progress — Python has no way to preempt a running sync frame — so it detects an overrun as soon as possible rather than truly interrupting one. Use tool_timeout to bound a single tool call instead:

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    tool_timeout=10.0,  # abandon any single tool call that runs past 10s
)
agent.add_tools(flaky_network_tool)

result = agent.invoke("Fetch the data and summarize it.")
# If flaky_network_tool hangs, its Observation becomes:
# "Error: tool 'flaky_network_tool' timed out after 10.0s."
# instead of blocking the agent loop forever.

A timed-out sync tool's underlying thread keeps running in the background (Python has no way to force-stop a running thread) — the agent loop itself just stops waiting on it. An async tool is actually cancelled at its next await point. tool_timeout=None (the default) disables this and matches prior behavior.

---

Scratchpad Size Limits

The scratchpad (agent.scratchpad, "prompt" mode only) is capped at Agent.MAX_SCRATCHPAD_CHARS (15,000 characters) by default — once exceeded, older steps are trimmed from the front and replaced with "[...earlier steps trimmed...]".

Character count alone is a poor proxy for what actually overflows an LLM's context window: tokens per character varies a lot by language and content (dense non-English text or code can run well under the ~4 chars/token rule of thumb, silently blowing a char-only budget's whole point long before 15,000 characters is reached). max_scratchpad_tokens adds a second, token-based cap on top of the character one:

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    max_scratchpad_tokens=4000,  # trim further if the scratchpad exceeds ~4000 tokens
)

Without a real tokenizer, token count is approximated as len(text) // 4 (the common English-prose rule of thumb). Pass token_counter= for precision — any fn(text: str) -> int, e.g.:

import tiktoken

encoding = tiktoken.encoding_for_model("gpt-4o")

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    max_scratchpad_tokens=4000,
    token_counter=lambda text: len(encoding.encode(text)),
)

max_scratchpad_tokens=None (the default) disables the token-based check — only the character cap applies, matching prior behavior.

Auto-Summarizing Scratchpad

The blunt trim above (drop older steps) throws away content. For long-running tool-heavy agents, built-in summarization periodically replaces agent.scratchpad with an LLM-generated summary instead, preserving key findings and tool results while shrinking the token footprint. It's implemented inline in the agent loop itself (not as middleware), turned on with three Agent() constructor kwargs:

from autourgos_agent import Agent

agent = Agent(llm=my_llm, summarize_every=5, max_scratchpad_chars=8000)

max_scratchpad_chars does double duty here — it's both the trim cap (Agent.MAX_SCRATCHPAD_CHARS) and the summarizer's own char-threshold trigger, sharing this one value. summarize_every=None (the default) leaves summarization disabled entirely, matching prior behavior.

Pass summarizer_llm= to use a separate, cheaper/faster model just for summarization instead of this agent's own llm (ignored if summarize_every isn't set):

from autourgos_openaichat import OpenAIChatModel

cheap_llm = OpenAIChatModel(model="gpt-4o-mini")
agent = Agent(llm=my_llm, summarize_every=5, summarizer_llm=cheap_llm)

Notes:

Parameter	Type	Default	Description
summarize_every	int | None	None	Summarize every N iterations. None disables summarization entirely.
max_scratchpad_chars	int	15000	Trim cap; also triggers summarization once the scratchpad exceeds it (when summarize_every is set).
summarizer_llm	any with .invoke()	None	Dedicated LLM for summarization; falls back to this agent's own llm when omitted.

• Has no effect in tool_calling_mode="native" — agent.scratchpad is a human-readable trace only in that mode and isn't sent to the LLM, so summarizing it wouldn't shrink the real context-window budget. Skips, warning once per agent.
• A concurrent summarization attempt for the same agent (e.g. two overlapping calls somehow racing) skips rather than blocking.
• On success, narrates via agent.logger.middleware("Summarizer", ...) (see Narrating middleware activity).

---

LLM Call Retries

By default, any failed LLM call (rate limit, network blip, transient 5xx) raises AgentLLMError immediately and ends the run — the same call would often succeed a moment later. llm_retries retries with exponential backoff instead:

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    llm_retries=3,             # retry up to 3 times before giving up
    llm_retry_backoff=1.0,     # base delay: 1s, 2s, 4s (capped below)
    llm_retry_max_backoff=30.0,
)
agent.add_tools(search_tool)

result = agent.invoke("What's the latest news?")
# A rate-limited call now retries instead of failing the whole run outright.

By default every exception is retried except NotImplementedError (the signal that tool_calling_mode="native" isn't supported by this LLM at all — a config error, not a transient one, so retrying it would just delay the clearer error). Pass llm_retry_on to customize which errors are worth retrying:

def only_rate_limits(exc: Exception) -> bool:
    return "rate limit" in str(exc).lower()

agent = Agent(llm=llm, llm_retries=5, llm_retry_on=only_rate_limits)

llm_retries=0 (the default) disables this and matches prior behavior — a single unconditional call, raising AgentLLMError on the first failure.

---

Custom System Prompt

Add extra instructions that persist across all steps.

agent = Agent(
    llm=OpenAIChatModel(model="gpt-4o"),
    system_prompt=(
        "You are a helpful financial analyst. "
        "Always cite your sources. "
        "Never speculate without data."
    ),
)
agent.add_tools(search_tool, calculator_tool)
result = agent.invoke("What is the P/E ratio of Apple?")

---

Constructor Reference

Parameter	Type	Default	Description
llm	any	None	LLM wrapper with .invoke() / .ainvoke(). Works with OpenAIChatModel, OpenAIResponse, or any compatible object
verbose	bool	False	Print step-by-step execution to stdout
full_output	bool	False	Also print raw LLM responses (implies verbose)
memory	MemoryProtocol	None	Memory backend for conversation history
max_iterations	int	15	Max Thought → Action → Observe cycles before stopping
max_execution_time	float	None	Wall-clock time limit in seconds
tool_timeout	float	None	Per-tool-call timeout in seconds. See Time and Iteration Limits
max_scratchpad_tokens	int	None	Extra token-based scratchpad budget on top of MAX_SCRATCHPAD_CHARS. See Scratchpad Size Limits
token_counter	callable	None	fn(text) -> int used to count tokens for max_scratchpad_tokens. Defaults to a len(text) // 4 approximation
llm_retries	int	0	Retries on a failed LLM call before raising AgentLLMError. See LLM Call Retries
llm_retry_backoff	float	1.0	Base delay in seconds between retries (exponential: backoff * 2**attempt)
llm_retry_max_backoff	float	30.0	Upper bound in seconds on the exponential backoff delay
llm_retry_on	callable	None	fn(exc) -> bool deciding whether a failure is worth retrying. Defaults to retrying everything except NotImplementedError
approval_callback	callable	None	Called as fn(tool_name, tool_input) before each tool. Return truthy to allow
middleware	list[CallbackHandler]	None	Event hooks for lifecycle events
max_consecutive_parse_errors	int	3	Stop after this many back-to-back JSON parse failures
tools	list[dict]	None	Initial tool list (more can be added with add_tools())
toolbox	list[Toolbox]	None	Toolboxes to lazy-load — hidden from the prompt until expose_toolbox()/expose_tool() is called. See Toolboxes
system_prompt	str	""	Extra system-level instruction added to every prompt
tool_calling_mode	"prompt" | "native"	"prompt"	"prompt": the original JSON-in-text agent loop. "native": uses the LLM's invoke_with_tools()/ainvoke_with_tools() — structured tool calls straight from the API, no JSON parsing, and multiple tool calls in one turn run concurrently. See Native Tool Calling
max_scratchpad_chars	int	None (class default 15,000)	Per-instance override of the scratchpad trim cap; also the built-in summarizer's char threshold when summarize_every is set. See Scratchpad Size Limits
summarize_every	int	None	Enables built-in scratchpad summarization every N iterations, using this agent's own llm (or summarizer_llm, if given). See Auto-Summarizing Scratchpad
summarizer_llm	any with .invoke()	None	Dedicated LLM the built-in summarizer uses instead of this agent's own llm. Only takes effect when summarize_every is also set
max_tool_output_chars	int	None (class default 5,000)	Per-instance override of the max characters kept from a single tool's result before truncating with "... [truncated]"
max_tool_workers	int	None (class default 8)	Per-instance override of the thread-pool size invoke() uses to run parallel tool calls. See Parallel Tool Calls
on_agent_start	callable	None	fn(query) (or fn(query, agent=self)) run every time this agent starts, without writing a full CallbackHandler. See on_agent_start Shortcut
history	str	None	Folder path — records every run to a Markdown + JSON file pair, with secrets redacted. See Run History
pre_iteration_callback	callable	None	Sync or async fn(iteration) run before every iteration. See Pre-Iteration Files & Callbacks
pre_iteration_files	str | list[str] | callable	None	File path(s) (or a callable returning them) injected into the LLM call at every iteration
image_quality	str | int	"auto"	Screenshot/image token-cost control for pre_iteration_files: "auto", "high", "medium", "low", or an int 1–100 JPEG quality. Ignored when pre_iteration_files isn't set

---

Tool Dict Reference

Key	Type	Required	Description
name	str	yes	Identifier used by the LLM. Use snake_case
description	str	yes	Plain-English description of what the tool does and when to use it
parameters	dict	recommended	JSON-Schema object describing the function's inputs
func	callable	yes	The Python function to call. Can be sync or async

parameters format (JSON Schema):

"parameters": {
    "type": "object",
    "properties": {
        "param_name": {
            "type": "string",       # string | number | integer | boolean | array | object
            "description": "...",   # shown to the LLM — make it clear
            "enum": ["a", "b"],     # optional: restrict to specific values
        },
    },
    "required": ["param_name"],     # list required params
}

---

What the Agent Returns

• Normal completion — invoke()/ainvoke() return a str: the final answer extracted from the LLM's final_answer field
• Error / limit reached — raises one of the exceptions below instead of returning a string, so callers catch a type rather than string-sniffing the result

---

Exceptions

All of these are exported from autourgos_agent and subclass AgentError:

Exception	Meaning
AgentTimeoutError	max_execution_time was exceeded
AgentMaxIterationsError	max_iterations reached without a final answer
AgentParseError	tool_calling_mode="prompt": LLM failed to produce valid JSON max_consecutive_parse_errors times in a row
AgentEmptyResponseError	tool_calling_mode="native": LLM returned neither a final answer nor tool calls max_consecutive_parse_errors times in a row
AgentLLMError	LLM raised an exception (network, rate limit, etc.) — the original exception is on .original

from autourgos_agent import Agent, AgentError, AgentTimeoutError

try:
    result = agent.invoke("...")
except AgentTimeoutError:
    ...
except AgentError:
    ...  # catches any of the above

---

v1 Backward Compatibility

The old Create_Agent factory function still works but emits a DeprecationWarning:

from autourgos_agent import Create_Agent  # DeprecationWarning

agent = Create_Agent(llm=llm)  # same as Agent(llm=llm)

Update your code to use Agent directly.

---

License

Apache License 2.0, Copyright (c) 2026 Jitin Kumar Sengar