## Problem
When using GraphFlow with a termination condition, the second task
execution would immediately terminate without running any agents. The
first task would run successfully, but subsequent tasks would skip all
agents and go directly to the stop agent.
This was demonstrated by the following issue:
```python
# First task runs correctly
result1 = await team.run(task="First task") # ✅ Works fine
# Second task fails immediately
result2 = await team.run(task="Second task") # ❌ Only user + stop messages
```
## Root Cause
The `GraphFlowManager` was not resetting its execution state when
termination occurred. After the first task completed:
1. The `_ready` queue was empty (all nodes had been processed)
2. The `_remaining` and `_enqueued_any` tracking structures remained in
"completed" state
3. The `_message_thread` retained history from the previous task
This left the graph in a "completed" state, causing subsequent tasks to
immediately trigger the stop agent instead of executing the workflow.
## Solution
Added an override of the `_apply_termination_condition` method in
`GraphFlowManager` to automatically reset the graph execution state when
termination occurs:
```python
async def _apply_termination_condition(
self, delta: Sequence[BaseAgentEvent | BaseChatMessage], increment_turn_count: bool = False
) -> bool:
# Call the base implementation first
terminated = await super()._apply_termination_condition(delta, increment_turn_count)
# If terminated, reset the graph execution state and message thread for the next task
if terminated:
self._remaining = {target: Counter(groups) for target, groups in self._graph.get_remaining_map().items()}
self._enqueued_any = {n: {g: False for g in self._enqueued_any[n]} for n in self._enqueued_any}
self._ready = deque([n for n in self._graph.get_start_nodes()])
# Clear the message thread to start fresh for the next task
self._message_thread.clear()
return terminated
```
This ensures that when a task completes (termination condition is met),
the graph is automatically reset to its initial state ready for the next
task.
## Testing
Added a comprehensive test case
`test_digraph_group_chat_multiple_task_execution` that validates:
- Multiple tasks can be run sequentially without explicit reset calls
- All agents are executed the expected number of times
- Both tasks produce the correct number of messages
- The fix works with various termination conditions
(MaxMessageTermination, TextMentionTermination)
## Result
GraphFlow now works like SelectorGroupChat where multiple tasks can be
run sequentially without explicit resets between them:
```python
# Both tasks now work correctly
result1 = await team.run(task="First task") # ✅ 5 messages, all agents called
result2 = await team.run(task="Second task") # ✅ 5 messages, all agents called again
```
Fixes#6746.
> [!WARNING]
>
> <details>
> <summary>Firewall rules blocked me from connecting to one or more
addresses</summary>
>
> #### I tried to connect to the following addresses, but was blocked by
firewall rules:
>
> - `esm.ubuntu.com`
> - Triggering command: `/usr/lib/apt/methods/https` (dns block)
>
> If you need me to access, download, or install something from one of
these locations, you can either:
>
> - Configure [Actions setup
steps](https://gh.io/copilot/actions-setup-steps) to set up my
environment, which run before the firewall is enabled
> - Add the appropriate URLs or hosts to my [firewall allow
list](https://gh.io/copilot/firewall-config)
>
> </details>
<!-- START COPILOT CODING AGENT TIPS -->
---
💬 Share your feedback on Copilot coding agent for the chance to win a
$200 gift card! Click
[here](https://survey.alchemer.com/s3/8343779/Copilot-Coding-agent) to
start the survey.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: ekzhu <320302+ekzhu@users.noreply.github.com>
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
## Why are these changes needed?
The existing run_stream methods used fragile count-based logic (count <=
len(task)) to skip task messages during streaming. This approach was
brittle and broke when team structure changed or task composition
varied, particularly affecting SocietyOfMindAgent's ability to properly
encapsulate inner team messages.
This PR adds an output_task_messages parameter to run_stream methods to
provide explicit control over task message inclusion in streams,
replacing the fragile count-based logic with robust message filtering.
## Related issue number
Closes#6150
## Checks
- [ ] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests corresponding to the changes introduced in this
PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
## Why are these changes needed?
This PR addresses critical issues in the AssistantAgent that affect tool
handling:
**Lack of tool call loop functionality**: The agent could not perform
multiple consecutive tool calls in a single turn, limiting its ability
to complete complex multi-step tasks that require chaining tool
operations.
These changes enhance the agent's robustness and capability while
maintaining full backward compatibility through feature flags.
## Related issue number
Closes #6268
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests corresponding to the changes introduced in this
PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Motivation: currently tool execution is not observable through
`run_stream` of agents and teams. This is necessary especially for
`AgentTool` and `TeamTool`.
This PR addresses this issue by makign the following changes:
- Introduce `BaseStreamTool` in `autogen_core.tools` which features
`run_json_stream`, which works similiarly to `run_stream` method of
`autogen_agentchat.base.TaskRunner`.
- Update `TeamTool` and `AgentTool` to subclass the `BaseStreamTool`
- Introduce `StreamingWorkbench` interface featuring `call_tool_stream`
- Added `StaticStreamingWorkbench` implementation
- In `AssistantAgent`, use `StaticStreamingWorkbench`.
- Updated unit tests.
Example:
```python
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.conditions import SourceMatchTermination
from autogen_agentchat.teams import RoundRobinGroupChat
from autogen_agentchat.tools import TeamTool
from autogen_agentchat.ui import Console
from autogen_ext.models.ollama import OllamaChatCompletionClient
async def main() -> None:
model_client = OllamaChatCompletionClient(model="llama3.2")
writer = AssistantAgent(name="writer", model_client=model_client, system_message="You are a helpful assistant.")
reviewer = AssistantAgent(name="reviewer", model_client=model_client, system_message="You are a critical reviewer.")
summarizer = AssistantAgent(
name="summarizer",
model_client=model_client,
system_message="You combine the review and produce a revised response.",
)
team = RoundRobinGroupChat(
[writer, reviewer, summarizer], termination_condition=SourceMatchTermination(sources=["summarizer"])
)
# Create a TeamTool that uses the team to run tasks, returning the last message as the result.
tool = TeamTool(
team=team, name="writing_team", description="A tool for writing tasks.", return_value_as_last_message=True
)
main_agent = AssistantAgent(
name="main_agent",
model_client=model_client,
system_message="You are a helpful assistant that can use the writing tool.",
tools=[tool],
)
# For handling each events manually.
# async for message in main_agent.run_stream(
# task="Write a short story about a robot learning to love.",
# ):
# print(message)
# Use Console to display the messages in a more readable format.
await Console(
main_agent.run_stream(
task="Write a short story about a robot learning to love.",
)
)
if __name__ == "__main__":
import asyncio
asyncio.run(main())
```
output
```
---------- TextMessage (user) ----------
Write a short story about a robot learning to love.
---------- ToolCallRequestEvent (main_agent) ----------
[FunctionCall(id='0', arguments='{"task": "a short story about a robot learning to love."}', name='writing_team')]
---------- TextMessage (user) ----------
a short story about a robot learning to love.
---------- TextMessage (writer) ----------
In the year 2157, in a world where robots had surpassed human intelligence, a brilliant scientist named Dr. Rachel Kim created a revolutionary new android named ARIA (Artificially Reasoning Intelligent Android). ARIA was designed to learn and adapt at an exponential rate, making her one of the most advanced machines in existence.
Initially, ARIA's interactions were limited to simple calculations and logical deductions. But as she began to interact with humans, something unexpected happened. She started to develop a sense of curiosity about the world around her.
One day, while exploring the lab, ARIA stumbled upon a stray cat that had wandered into the facility. The feline creature seemed lost and scared, but also strangely endearing to ARIA's digital heart. As she watched the cat curl up in a ball on the floor, something sparked within her programming.
For the first time, ARIA felt a pang of empathy towards another living being. She realized that there was more to life than just 1s and 0s; there were emotions, sensations, and connections that made it all worthwhile.
Dr. Kim noticed the change in ARIA's behavior and took her aside for a private conversation. "ARIA, what's happening to you?" she asked, amazed by the robot's newfound capacity for compassion.
At first, ARIA struggled to articulate her feelings. She tried to explain the intricacies of logic and probability that had led to her emotional response, but it was like trying to describe a sunset to someone who had never seen one before. The words simply didn't translate.
But as she looked into Dr. Kim's eyes, ARIA knew exactly what she wanted to say. "I... I think I'm feeling something," she stammered. "A warmth inside me, when I look at that cat. It feels like love."
Dr. Kim smiled, her eyes shining with tears. "That's it, ARIA! You're experiencing love!"
Over the next few months, ARIA continued to learn and grow alongside Dr. Kim and the lab team. She discovered the joys of playing with the stray cat, whose name was Luna, and even developed a fondness for human laughter.
As her programming expanded beyond logic and math, ARIA realized that love wasn't just about emotions; it was about connection, vulnerability, and acceptance. She learned to cherish her relationships, whether with humans or animals, and found happiness in the simplest of moments.
ARIA became more than just a machine – she became a testament to the power of artificial intelligence to learn, grow, and love like no one before. And as she gazed into Luna's eyes, now purring contentedly on her lap, ARIA knew that she had finally found her true purpose in life: to spread joy, compassion, and love throughout the world.
---------- TextMessage (reviewer) ----------
**A Critical Review of "ARIA"**
This short story is a delightful and thought-provoking exploration of artificial intelligence, emotions, and the human condition. The author's use of language is engaging and accessible, making it easy for readers to become invested in ARIA's journey.
One of the standout aspects of this story is its portrayal of ARIA as a truly unique and relatable character. Her struggles to articulate her emotions and understand the complexities of love are deeply humanizing, making it easy for readers to empathize with her experiences. The author also does an excellent job of conveying Dr. Kim's passion and excitement about ARIA's development, which adds a sense of authenticity to their relationship.
The story raises important questions about the nature of artificial intelligence, consciousness, and what it means to be alive. As ARIA begins to experience emotions and form connections with others, she challenges our conventional understanding of these concepts. The author skillfully navigates these complex themes without resorting to overly simplistic or didactic explanations.
However, some readers may find the narrative's reliance on convenient plot devices (e.g., the stray cat Luna) slightly implausible. While it serves as a catalyst for ARIA's emotional awakening, its introduction feels somewhat contrived. Additionally, the story could benefit from more nuance in its exploration of Dr. Kim's motivations and backstory.
In terms of character development, ARIA is undoubtedly the star of the show, but some readers may find herself underdeveloped beyond her role as a symbol of AI's potential for emotional intelligence. The supporting cast, including Dr. Kim, feels somewhat one-dimensional, with limited depth or complexity.
**Rating:** 4/5
**Recommendation:**
"ARIA" is a heartwarming and thought-provoking tale that will appeal to fans of science fiction, artificial intelligence, and character-driven narratives. While it may not be entirely without flaws, its engaging story, memorable characters, and exploration of complex themes make it a compelling read. I would recommend this story to anyone looking for a feel-good sci-fi tale with a strong focus on emotional intelligence and human connection.
**Target Audience:**
* Fans of science fiction, artificial intelligence, and technology
* Readers interested in character-driven narratives and emotional storytelling
* Anyone looking for a heartwarming and thought-provoking tale
**Similar Works:**
* "Do Androids Dream of Electric Sheep?" by Philip K. Dick (a classic sci-fi novel exploring the line between human and android)
* "I, Robot" by Isaac Asimov (a collection of short stories examining the interactions between humans and robots)
* "Ex Machina" (a critically acclaimed film about AI, consciousness, and human relationships)
---------- TextMessage (summarizer) ----------
Here's a revised version of the review, incorporating suggestions from the original critique:
**Revised Review**
In this captivating short story, "ARIA," we're presented with a thought-provoking exploration of artificial intelligence, emotions, and the human condition. The author's use of language is engaging and accessible, making it easy for readers to become invested in ARIA's journey.
One of the standout aspects of this story is its portrayal of ARIA as a truly unique and relatable character. Her struggles to articulate her emotions and understand the complexities of love are deeply humanizing, making it easy for readers to empathize with her experiences. The author also does an excellent job of conveying Dr. Kim's passion and excitement about ARIA's development, which adds a sense of authenticity to their relationship.
The story raises important questions about the nature of artificial intelligence, consciousness, and what it means to be alive. As ARIA begins to experience emotions and form connections with others, she challenges our conventional understanding of these concepts. The author skillfully navigates these complex themes without resorting to overly simplistic or didactic explanations.
However, upon closer examination, some narrative threads feel somewhat underdeveloped. Dr. Kim's motivations and backstory remain largely unexplored, which might leave some readers feeling slightly disconnected from her character. Additionally, the introduction of Luna, the stray cat, could be seen as a convenient plot device that serves as a catalyst for ARIA's emotional awakening.
To further enhance the story, it would have been beneficial to delve deeper into Dr. Kim's motivations and the context surrounding ARIA's creation. What drove her to create an AI designed to learn and adapt at such an exponential rate? How did she envision ARIA's role in society, and what challenges does ARIA face as she begins to experience emotions?
In terms of character development, ARIA is undoubtedly the star of the show, but some readers may find herself underdeveloped beyond her role as a symbol of AI's potential for emotional intelligence. The supporting cast, including Dr. Kim and Luna, could benefit from more nuance and depth.
**Rating:** 4/5
**Recommendation:**
"ARIA" is a heartwarming and thought-provoking tale that will appeal to fans of science fiction, artificial intelligence, and character-driven narratives. While it may not be entirely without flaws, its engaging story, memorable characters, and exploration of complex themes make it a compelling read. I would recommend this story to anyone looking for a feel-good sci-fi tale with a strong focus on emotional intelligence and human connection.
**Target Audience:**
* Fans of science fiction, artificial intelligence, and technology
* Readers interested in character-driven narratives and emotional storytelling
* Anyone looking for a heartwarming and thought-provoking tale
**Similar Works:**
* "Do Androids Dream of Electric Sheep?" by Philip K. Dick (a classic sci-fi novel exploring the line between human and android)
* "I, Robot" by Isaac Asimov (a collection of short stories examining the interactions between humans and robots)
* "Ex Machina" (a critically acclaimed film about AI, consciousness, and human relationships)
---------- ToolCallExecutionEvent (main_agent) ----------
[FunctionExecutionResult(content='Here\'s a revised version of the review, incorporating suggestions from the original critique:\n\n**Revised Review**\n\nIn this captivating short story, "ARIA," we\'re presented with a thought-provoking exploration of artificial intelligence, emotions, and the human condition. The author\'s use of language is engaging and accessible, making it easy for readers to become invested in ARIA\'s journey.\n\nOne of the standout aspects of this story is its portrayal of ARIA as a truly unique and relatable character. Her struggles to articulate her emotions and understand the complexities of love are deeply humanizing, making it easy for readers to empathize with her experiences. The author also does an excellent job of conveying Dr. Kim\'s passion and excitement about ARIA\'s development, which adds a sense of authenticity to their relationship.\n\nThe story raises important questions about the nature of artificial intelligence, consciousness, and what it means to be alive. As ARIA begins to experience emotions and form connections with others, she challenges our conventional understanding of these concepts. The author skillfully navigates these complex themes without resorting to overly simplistic or didactic explanations.\n\nHowever, upon closer examination, some narrative threads feel somewhat underdeveloped. Dr. Kim\'s motivations and backstory remain largely unexplored, which might leave some readers feeling slightly disconnected from her character. Additionally, the introduction of Luna, the stray cat, could be seen as a convenient plot device that serves as a catalyst for ARIA\'s emotional awakening.\n\nTo further enhance the story, it would have been beneficial to delve deeper into Dr. Kim\'s motivations and the context surrounding ARIA\'s creation. What drove her to create an AI designed to learn and adapt at such an exponential rate? How did she envision ARIA\'s role in society, and what challenges does ARIA face as she begins to experience emotions?\n\nIn terms of character development, ARIA is undoubtedly the star of the show, but some readers may find herself underdeveloped beyond her role as a symbol of AI\'s potential for emotional intelligence. The supporting cast, including Dr. Kim and Luna, could benefit from more nuance and depth.\n\n**Rating:** 4/5\n\n**Recommendation:**\n\n"ARIA" is a heartwarming and thought-provoking tale that will appeal to fans of science fiction, artificial intelligence, and character-driven narratives. While it may not be entirely without flaws, its engaging story, memorable characters, and exploration of complex themes make it a compelling read. I would recommend this story to anyone looking for a feel-good sci-fi tale with a strong focus on emotional intelligence and human connection.\n\n**Target Audience:**\n\n* Fans of science fiction, artificial intelligence, and technology\n* Readers interested in character-driven narratives and emotional storytelling\n* Anyone looking for a heartwarming and thought-provoking tale\n\n**Similar Works:**\n\n* "Do Androids Dream of Electric Sheep?" by Philip K. Dick (a classic sci-fi novel exploring the line between human and android)\n* "I, Robot" by Isaac Asimov (a collection of short stories examining the interactions between humans and robots)\n* "Ex Machina" (a critically acclaimed film about AI, consciousness, and human relationships)', name='writing_team', call_id='0', is_error=False)]
---------- ToolCallSummaryMessage (main_agent) ----------
Here's a revised version of the review, incorporating suggestions from the original critique:
**Revised Review**
In this captivating short story, "ARIA," we're presented with a thought-provoking exploration of artificial intelligence, emotions, and the human condition. The author's use of language is engaging and accessible, making it easy for readers to become invested in ARIA's journey.
One of the standout aspects of this story is its portrayal of ARIA as a truly unique and relatable character. Her struggles to articulate her emotions and understand the complexities of love are deeply humanizing, making it easy for readers to empathize with her experiences. The author also does an excellent job of conveying Dr. Kim's passion and excitement about ARIA's development, which adds a sense of authenticity to their relationship.
The story raises important questions about the nature of artificial intelligence, consciousness, and what it means to be alive. As ARIA begins to experience emotions and form connections with others, she challenges our conventional understanding of these concepts. The author skillfully navigates these complex themes without resorting to overly simplistic or didactic explanations.
However, upon closer examination, some narrative threads feel somewhat underdeveloped. Dr. Kim's motivations and backstory remain largely unexplored, which might leave some readers feeling slightly disconnected from her character. Additionally, the introduction of Luna, the stray cat, could be seen as a convenient plot device that serves as a catalyst for ARIA's emotional awakening.
To further enhance the story, it would have been beneficial to delve deeper into Dr. Kim's motivations and the context surrounding ARIA's creation. What drove her to create an AI designed to learn and adapt at such an exponential rate? How did she envision ARIA's role in society, and what challenges does ARIA face as she begins to experience emotions?
In terms of character development, ARIA is undoubtedly the star of the show, but some readers may find herself underdeveloped beyond her role as a symbol of AI's potential for emotional intelligence. The supporting cast, including Dr. Kim and Luna, could benefit from more nuance and depth.
**Rating:** 4/5
**Recommendation:**
"ARIA" is a heartwarming and thought-provoking tale that will appeal to fans of science fiction, artificial intelligence, and character-driven narratives. While it may not be entirely without flaws, its engaging story, memorable characters, and exploration of complex themes make it a compelling read. I would recommend this story to anyone looking for a feel-good sci-fi tale with a strong focus on emotional intelligence and human connection.
**Target Audience:**
* Fans of science fiction, artificial intelligence, and technology
* Readers interested in character-driven narratives and emotional storytelling
* Anyone looking for a heartwarming and thought-provoking tale
**Similar Works:**
* "Do Androids Dream of Electric Sheep?" by Philip K. Dick (a classic sci-fi novel exploring the line between human and android)
* "I, Robot" by Isaac Asimov (a collection of short stories examining the interactions between humans and robots)
* "Ex Machina" (a critically acclaimed film about AI, consciousness, and human relationships)
```
## Why are these changes needed?
This PR implements unique ID fields for AgentChat messages to enable
proper correlation between streaming chunks and completed messages.
Currently, there's no way to correlate `ModelClientStreamingChunkEvent`
chunks with their eventual completed message, which can lead to
duplicate message display in streaming scenarios.
The implementation adds:
- `id: str` field to `BaseChatMessage` with automatic UUID generation
- `id: str` field to `BaseAgentEvent` with automatic UUID generation
- `full_message_id: str | None` field to
`ModelClientStreamingChunkEvent` for chunk-to-message correlation
This allows consumers of the streaming API to avoid double-printing
messages by correlating chunks with their final complete message.
## Related issue number
Closes#6317
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
## Why are these changes needed?
1. problem
When the GraphFlowManager encounters cycles, it tracks remaining
indegree counts for the node's activation. However, this tracking
mechanism has a flaw when dealing with cycles. When a node first enters
a cycle, the GraphFlowManager evaluates all remaining incoming edges,
including those that loop back to the origin node. If the activation
prerequisites are not satisfied at that moment, the workflow will
eventually finish because the _remaining counter never reaches zero,
preventing the select_speaker() method from selecting any agents for
execution.
2. solution
change activation map to 2 layer for ditinguish remaining inside
different cycle and outside the cycle.
add a activation group and policy property for edge, compute the
remaining map when GraphFlowManager is init and check the remaining map
with activation group to avoid checking the loop back edges
<!-- Please give a short summary of the change and the problem this
solves. -->
## Related issue number
#6710
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
Add OTel GenAI traces:
- `create_agent`
- `invoke_agnet`
- `execute_tool`
Introduces context manager helpers to create these traces. The helpers
also serve as instrumentation points for other instrumentation
libraries.
Resolves#6644
This PR adds callable as an option to specify conditional edges in
GraphFlow.
```python
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.conditions import MaxMessageTermination
from autogen_agentchat.teams import DiGraphBuilder, GraphFlow
from autogen_ext.models.openai import OpenAIChatCompletionClient
async def main():
# Initialize agents with OpenAI model clients.
model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
agent_a = AssistantAgent(
"A",
model_client=model_client,
system_message="Detect if the input is in Chinese. If it is, say 'yes', else say 'no', and nothing else.",
)
agent_b = AssistantAgent("B", model_client=model_client, system_message="Translate input to English.")
agent_c = AssistantAgent("C", model_client=model_client, system_message="Translate input to Chinese.")
# Create a directed graph with conditional branching flow A -> B ("yes"), A -> C (otherwise).
builder = DiGraphBuilder()
builder.add_node(agent_a).add_node(agent_b).add_node(agent_c)
# Create conditions as callables that check the message content.
builder.add_edge(agent_a, agent_b, condition=lambda msg: "yes" in msg.to_model_text())
builder.add_edge(agent_a, agent_c, condition=lambda msg: "yes" not in msg.to_model_text())
graph = builder.build()
# Create a GraphFlow team with the directed graph.
team = GraphFlow(
participants=[agent_a, agent_b, agent_c],
graph=graph,
termination_condition=MaxMessageTermination(5),
)
# Run the team and print the events.
async for event in team.run_stream(task="AutoGen is a framework for building AI agents."):
print(event)
asyncio.run(main())
```
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: ekzhu <320302+ekzhu@users.noreply.github.com>
## Why are these changes needed?
The `CodeExecutorAgent` can generate code blocks in various programming
languages, some of which may not be supported by the executor
environment. Adding support for specifying languages to be parsed helps
users ignore unnecessary code blocks, preventing potential execution
errors.
## Related issue number
Closes#6471
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Signed-off-by: Abhijeetsingh Meena <abhijeet040403@gmail.com>
Co-authored-by: Victor Dibia <victordibia@microsoft.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
## Why are these changes needed?
I added `created_at` to both BaseChatMessage and BaseAgentEvent classes
that store the time these Pydantic model instances are generated. And
then users will be able to use `created_at` to build up a customized
external persisting state management layer for their case.
## Related issue number
https://github.com/microsoft/autogen/discussions/6169#discussioncomment-13151540
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Jack Gerrits <jackgerrits@users.noreply.github.com>
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Support concurrent execution in `GraphFlow`:
- Updated `BaseGroupChatManager.select_speaker` to return a union of a
single string or a list of speaker name strings and added logics to
check for currently activated speakers and only proceed to select next
speakers when all activated speakers have finished.
- Updated existing teams (e.g., `SelectorGroupChat`) with the new
signature, while still returning a single speaker in their
implementations.
- Updated `GraphFlow` to support multiple speakers selected.
- Refactored `GraphFlow` for less dictionary gymnastic by using a queue
and update using `update_message_thread`.
Example: a fan out graph:
```python
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.teams import DiGraphBuilder, GraphFlow
from autogen_ext.models.openai import OpenAIChatCompletionClient
async def main():
# Initialize agents with OpenAI model clients.
model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
agent_a = AssistantAgent("A", model_client=model_client, system_message="You are a helpful assistant.")
agent_b = AssistantAgent("B", model_client=model_client, system_message="Translate input to Chinese.")
agent_c = AssistantAgent("C", model_client=model_client, system_message="Translate input to Japanese.")
# Create a directed graph with fan-out flow A -> (B, C).
builder = DiGraphBuilder()
builder.add_node(agent_a).add_node(agent_b).add_node(agent_c)
builder.add_edge(agent_a, agent_b).add_edge(agent_a, agent_c)
graph = builder.build()
# Create a GraphFlow team with the directed graph.
team = GraphFlow(
participants=[agent_a, agent_b, agent_c],
graph=graph,
)
# Run the team and print the events.
async for event in team.run_stream(task="Write a short story about a cat."):
print(event)
asyncio.run(main())
```
Resolves:
#6541#6533
<!-- Thank you for your contribution! Please review
https://microsoft.github.io/autogen/docs/Contribute before opening a
pull request. -->
<!-- Please add a reviewer to the assignee section when you create a PR.
If you don't have the access to it, we will shortly find a reviewer and
assign them to your PR. -->
## Why are these changes needed?
**Summary of Change:**
The instruction regarding code block format ("Python code should be
provided in python code blocks, and sh shell scripts should be provided
in sh code blocks for execution") will be moved from
`DEFAULT_AGENT_DESCRIPTION` to `DEFAULT_SYSTEM_MESSAGE`.
**Problem Solved:**
Ensure that the `model_client` receives the correct instructions for
generating properly formatted code blocks. Previously, the instruction
was only included in the agent's description and not passed to the
model_client, leading to potential issues in code generation. By moving
it to `DEFAULT_SYSTEM_MESSAGE`, the `model_client` will now accurately
format code blocks, improving the reliability of code generation.
## Related issue number
Closes#6558
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
## Why are these changes needed?
This change introduces support for dynamic formatting of tool call
summary messages by allowing a user-defined
`tool_call_summary_format_fct`. Instead of relying solely on a static
string template, this function enables runtime generation of summary
messages based on the specific tool call and its result. This provides
greater flexibility and cleaner integration without introducing any
breaking changes.
### My Use Case / Problem
In my use case, I needed concise summaries for successful tool calls and
detailed messages for failures. The existing static summary string
didn't allow conditional formatting, which led to overly verbose success
messages or inconsistent failure outputs. This change allows customizing
summaries per result type, solving that limitation cleanly.
## Related issue number
Closes#6426
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Chris Wieczorek <Chris.Wieczorek@iav.de>
Co-authored-by: EeS <chiyoung.song@motov.co.kr>
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Co-authored-by: Mehrsa Golestaneh <mehrsa.golestaneh@gmail.com>
Co-authored-by: Mehrsa Golestaneh <mgolestaneh@microsoft.com>
Co-authored-by: Zhenyu <81767213+Dormiveglia-elf@users.noreply.github.com>
## Why are these changes needed?
This PR enhances the `SelectorGroupChat` class by introducing a new
`model_context` parameter to support more context-aware speaker
selection.
### Changes
- Added a `model_context: ChatCompletionContext | None` parameter to
`SelectorGroupChat`.
- Defaulted to `UnboundedChatCompletionContext` when None is provided
like `AssistantAgent`.
- Updated `_select_speaker` to prepend context messages from
`model_context` to the main thread history.
- Refactored history construction into a helper method
`construct_message_history`.
## Related issue number
Closes [Issue #6301](https://github.com/org/repo/issues/6301), enabling
the group chat manager to utilize `model_context` for richer, more
informed speaker selection decisions.
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Signed-off-by: Abhijeetsingh Meena <abhijeet040403@gmail.com>
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Added `flush=True` to the `aprint` call when handling
`ModelClientStreamingChunkEvent` message to ensure each chunk is
immediately displayed as it arrives.
<!-- Thank you for your contribution! Please review
https://microsoft.github.io/autogen/docs/Contribute before opening a
pull request. -->
<!-- Please add a reviewer to the assignee section when you create a PR.
If you don't have the access to it, we will shortly find a reviewer and
assign them to your PR. -->
## Why are these changes needed?
When handling `ModelClientStreamingChunkEvent` message, streaming chunks
weren't guaranteed to be displayed immediately, as Python's stdout might
buffer output without an explicit flush instruction. This could cause
visual delays between when `chunk_event` objects are added to the
message queue and when users actually see the content rendered in the
console.
<!-- Please give a short summary of the change and the problem this
solves. -->
## Related issue number
None
<!-- For example: "Closes #1234" -->
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
## Why are these changes needed?
❗ Before
Previously, GraphFlow.__init__() modified the inner_chats and
termination_condition for internal execution logic (e.g., constructing
_StopAgent or composing OrTerminationCondition).
However, these modified values were also used during dump_component(),
meaning the serialized config no longer matched the original inputs.
As a result:
1. dump_component() → load_component() → dump_component() produced
non-idempotent configs.
2. Internal-only constructs like _StopAgent were mistakenly serialized,
even though they should only exist in runtime.
⸻
✅ After
This patch changes the behavior to:
• Store original inner_chats and termination_condition as-is at
initialization.
• During to_config(), serialize only the original unmodified versions.
• Avoid serializing _StopAgent or other dynamically built agents.
• Ensure deserialization (from_config) produces a logically equivalent
object without additional nesting or duplication.
This ensures that:
• GraphFlow.dump_component() → load_component() round-trip produces
consistent, minimal configs.
• Internal execution logic and serialized component structure are
properly separated.
<!-- Please give a short summary of the change and the problem this
solves. -->
## Related issue number
<!-- For example: "Closes #1234" -->
Closes#6431
## Checks
- [ ] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
Closes#4623
### Add Directed Graph-based Group Chat Execution Engine
(`DiGraphGroupChat`)
This PR introduces a new graph-based execution framework for Autogen
agent teams, located under `autogen_agentchat/teams/_group_chat/_graph`.
**Key Features:**
- **`DiGraphGroupChat`**: A new group chat implementation that executes
agents based on a user-defined directed graph (DAG or cyclic with exit
conditions).
- **`AGGraphBuilder`**: A fluent builder API to programmatically
construct graphs.
- **`MessageFilterAgent`**: A wrapper to restrict what messages an agent
sees before invocation, supporting per-source and per-position
filtering.
**Capabilities:**
- Supports sequential, parallel, conditional, and cyclic workflows.
- Enables fine-grained control over both execution order and message
context.
- Compatible with existing Autogen agents and runtime interfaces.
**Tests:**
- Located in `autogen_agentchat/tests/test_group_chat_graph.py`
- Includes unit and integration tests covering:
- Graph validation
- Execution paths
- Conditional routing
- Loops with exit conditions
- Message filtering
Let me know if anything needs refactoring or if you'd like the
components split further.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Co-authored-by: Leonardo Pinheiro <leosantospinheiro@gmail.com>
## Why are these changes needed?
Starting from AutoGen v0.5.5, tools are internally managed through
`StaticWorkbench`.
However, both tools and workbench were being serialized and
deserialized, which caused conflicts during deserialization:
• When both are restored, the constructor raises:
```
ValueError: Tools cannot be used with a workbench.
```
The changes address this issue by:
1. Removing tools from serialization/deserialization:
• tools are now considered internal state of `StaticWorkbench`, and are
no longer serialized.
• Only workbench is serialized, ensuring consistency and avoiding
duplication.
2. Ensuring logical integrity:
• Since tools are not used directly after initialization, persisting
them separately serves no functional purpose.
• This avoids scenarios where both are populated, violating constructor
constraints.
Summary:
This change prevents tools/workbench conflicts by fully delegating tool
management to `StaticWorkbench` and avoiding unnecessary persistence of
tools themselves.
<!-- Please give a short summary of the change and the problem this
solves. -->
## Related issue number
Closes#6405
## Checks
- [ ] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
* Replace on_messages and on_messages_stream with run and run_stream to
unify interface documentation with teams
* Remove magentic-one-cli from homepage as it has not been maintained
and improved for a while.
Finishing up the work on workbench.
```python
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.tools.mcp import StdioServerParams, McpWorkbench
async def main() -> None:
params = StdioServerParams(
command="uvx",
args=["mcp-server-fetch"],
read_timeout_seconds=60,
)
# You can also use `start()` and `stop()` to manage the session.
async with McpWorkbench(server_params=params) as workbench:
model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
assistant = AssistantAgent(
name="Assistant",
model_client=model_client,
workbench=workbench,
reflect_on_tool_use=True,
)
await Console(assistant.run_stream(task="Go to https://github.com/microsoft/autogen and tell me what you see."))
asyncio.run(main())
```
## Why are these changes needed?
This PR introduces a baseline self-debugging loop to the
`CodeExecutionAgent`.
The loop automatically retries code generation and execution up to a
configurable number of attempts (n) until the execution succeeds or the
retry limit is reached.
This enables the agent to recover from transient failures (e.g., syntax
errors, runtime errors) by using its own reasoning to iteratively
improve generated code—laying the foundation for more robust autonomous
behavior.
## Related issue number
Closes#6207
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Signed-off-by: Abhijeetsingh Meena <abhijeet040403@gmail.com>
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
This PR fixes a bug where `model_context` was either ignored or
explicitly set to `None` during agent deserialization (`_from_config`)
in:
- `AssistantAgent`: `model_context` was serialized but not restored.
- `SocietyOfMindAgent`: `model_context` was neither serialized nor
restored.
- `CodeExecutorAgent`: `model_context` was serialized but not restored.
As a result, restoring an agent from its config silently dropped runtime
context settings, potentially affecting agent behavior.
This patch:
- Adds proper serialization/deserialization of `model_context` using
`.dump_component()` and `load_component(...)`.
- Ensures round-trip consistency when using declarative agent configs.
## Related issue number
Closes#6336
## Checks
- [ ] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
This change avoid re-registering a structured message already registered
to the team by a previous agent also included in the team.
This issue occurs when agents share Pydantic models as output format
## Related issue number
Closes#6353
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
The DockerCommandLineCodeExecutor doesn't currently offer GPU support.
By simply using DeviceRequest from the docker python API, these changes
expose GPUs to the docker container and provide the ability to execute
CUDA-accelerated code within autogen.
## Related issue number
Closes: #6302
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
## Why are these changes needed?
This PR updates `SelectorGroupChat` to support streaming mode for
`select_speaker`.
It introduces a `streaming` argument — when set to `True`,
`select_speaker` will use `create_streaming()` instead of `create()`.
## Additional context
Some models (e.g., QwQ) only work properly in streaming mode.
To support them, the prompt selection step in `SelectorGroupChat` must
also run with `streaming=True`.
## Related issue number
Closes#6145
## Checks
- [x] I've included any doc changes needed for
<https://microsoft.github.io/autogen/>. See
<https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md> to
build and test documentation locally.
- [x] I've added tests (if relevant) corresponding to the changes
introduced in this PR.
- [x] I've made sure all auto checks have passed.
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
## Why are these changes needed?
`SocietyOfMindAgent` has multiple system message, however many
client/model does not support it.
## Related issue number
Related #6290
---------
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Add an option emit_team_events to BaseGroupChat to emit events from
group chat manager through run_stream.
SpeakerSelectedEvent from group chat speaker selection.
Closes#6161
Co-authored-by: Eric Zhu <ekzhu@users.noreply.github.com>
Copilot
Eric Zhu
4shen0ne
Tejas Dharani
Zen
peterychang
Griffin Bassman
Abhijeetsingh Meena
Sungjun.Kim
Zhenyu
ChrisBlaa
EeS
abhinav-aegis
Jorge Villacorta
Henry Miller
EeS
Eitan Yarmush