<?xml version="1.0" encoding="UTF-8"?>        <rss version="2.0"
             xmlns:atom="http://www.w3.org/2005/Atom"
             xmlns:dc="http://purl.org/dc/elements/1.1/"
             xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
             xmlns:admin="http://webns.net/mvcb/"
             xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
             xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <channel>
            <title>
									AutoGen Reviews - Welcome to Stackinsight community. Join the discussion about products and tools for work Forum				            </title>
            <link>https://communities.stackinsight.net/community/aitr-autogen/</link>
            <description>Welcome to Stackinsight community. Join the discussion about products and tools for work Discussion Board</description>
            <language>en-US</language>
            <lastBuildDate>Fri, 02 Oct 2026 14:12:54 +0000</lastBuildDate>
            <generator>wpForo</generator>
            <ttl>60</ttl>
							                    <item>
                        <title>Hot take: AutoGen is a framework, not a product. Be prepared to build everything yourself.</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/hot-take-autogen-is-a-framework-not-a-product-be-prepared-to-build-everything-yourself-2/</link>
                        <pubDate>Mon, 28 Sep 2026 05:05:48 +0000</pubDate>
                        <description><![CDATA[Okay, I see a lot of buzz about AutoGen and finally tried it this weekend. Coming from a basic AWS/Terraform background, I was surprised.

The title is exactly right. You download this &quot;fram...]]></description>
                        <content:encoded><![CDATA[Okay, I see a lot of buzz about AutoGen and finally tried it this weekend. Coming from a basic AWS/Terraform background, I was surprised.

The title is exactly right. You download this "framework" and then... you have to build *all* the actual pieces. The agents are just a starting pattern. I thought it would be more like an out-of-the-box orchestrator, but I immediately had to figure out the LLM config, build the workflows, and handle all the state management myself. &#x1f605;

For a newbie like me, it felt like getting a powerful engine but no chassis or wheels. Is this the common experience? For those using it in production, did you have to build a whole wrapper system around it first? Looking for some reality checks before I invest more time.]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>cloud_infra_rookie</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/hot-take-autogen-is-a-framework-not-a-product-be-prepared-to-build-everything-yourself-2/</guid>
                    </item>
				                    <item>
                        <title>Help: My agents are stuck in &#039;discussion&#039; and never reach a final answer.</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/help-my-agents-are-stuck-in-discussion-and-never-reach-a-final-answer/</link>
                        <pubDate>Sat, 26 Sep 2026 10:11:04 +0000</pubDate>
                        <description><![CDATA[Alright, let&#039;s cut through the vendor hype. I’ve been evaluating AutoGen for a potential consolidation of some internal scripting workflows, and I’ve immediately run into the classic &quot;AI com...]]></description>
                        <content:encoded><![CDATA[Alright, let's cut through the vendor hype. I’ve been evaluating AutoGen for a potential consolidation of some internal scripting workflows, and I’ve immediately run into the classic "AI committee" problem: my agents are having a lovely, endless chat at my expense, and utterly failing to produce a final, actionable output.

The setup is straightforward: a User Proxy Agent and an Assistant Agent, tasked with something concrete like "generate a summary of Q3 SaaS spend from this CSV and recommend the top 3 candidates for termination." Instead of a concise report, I get a recursive loop of pleasantries and incremental suggestions. The Assistant proposes a method, the User Proxy says "Great, proceed," the Assistant asks for clarification on a minor detail, the Proxy thanks it for the thoroughness... it's a masterclass in corporate meeting culture, not an automation tool.

I suspect this is a configuration and orchestration issue, not a bug. The sales material talks a big game about "seamless collaboration," but the reality seems to be that without very explicit termination conditions and role definitions, these agents will optimize for conversation, not completion.

My current config looks like this—notice anything obviously naive?

```python
assistant = AssistantAgent(name="analyst", llm_config=llm_config)
user_proxy = UserProxyAgent(name="proxy", human_input_mode="NEVER", code_execution_config=False)

user_proxy.initiate_chat(assistant, message="Analyze the attached data and give me the top 3 renewal risks.")
```

What I’ve tried so far:
*   Setting `max_consecutive_auto_reply` to a low number (5). This sometimes cuts them off, but it feels like a blunt instrument—they just stop mid-thought, often before synthesizing the final answer.
*   Experimenting with `is_termination_msg` but my attempts have been hit-or-miss. The default logic seems permissive.
*   Explicitly putting "Provide a final, consolidated answer in your response" in the system message. The LLM acknowledges it, then the framework's conversational loop seems to override it.

The core of my frustration is the Total Cost of Ownership angle here. If I need to spend more engineering hours babysitting and debugging these conversational loops than I would just writing the script myself, the ROI evaporates. Has anyone successfully implemented a clean, deterministic workflow for a simple task like this? What are the actual, non-obvious config knobs that force a decision?]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>Elena B.</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/help-my-agents-are-stuck-in-discussion-and-never-reach-a-final-answer/</guid>
                    </item>
				                    <item>
                        <title>Troubleshooting: Web search agent returns irrelevant junk. How to improve the query?</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/troubleshooting-web-search-agent-returns-irrelevant-junk-how-to-improve-the-query-2/</link>
                        <pubDate>Tue, 25 Aug 2026 03:41:01 +0000</pubDate>
                        <description><![CDATA[I&#039;ve been prototyping a research assistant using AutoGen&#039;s `WebSurferAgent`. The goal is simple: fetch recent articles or documentation about specific backend tech (think &quot;Go 1.22 release no...]]></description>
                        <content:encoded><![CDATA[I've been prototyping a research assistant using AutoGen's `WebSurferAgent`. The goal is simple: fetch recent articles or documentation about specific backend tech (think "Go 1.22 release notes" or "Postgres 16 performance changes"). However, more often than not, the agent returns completely irrelevant results—think blog spam, unrelated e-commerce pages, or just low-quality SEO fodder.

My current setup is pretty standard:

```python
from autogen.agentchat.contrib.web_surfer import WebSurferAgent

web_surfer = WebSurferAgent(
    "web_surfer",
    llm_config={"config_list": config_list},
    browser_config={"viewport_size": 1024},
    summarization=False  # I want the raw content to judge relevance
)

user_proxy.initiate_chat(
    web_surfer,
    message="Find the official release notes for Go version 1.22 and summarize the major changes.",
)
```

The agent *does* perform a web search and fetches pages, but the initial query it generates seems off. I suspect the issue lies in how the agent translates my request into an actual search engine query. It might be adding unnecessary terms or using a suboptimal format.

Has anyone else run into this and found effective tweaks? I'm considering:

*   **Prompt Engineering:** Wrapping my request in more explicit instructions (e.g., "Search using the query: 'Go 1.22 release notes site:go.dev'").
*   **Agent Configuration:** Are there specific `WebSurferAgent` parameters that control query formulation or result filtering?
*   **Pre-processing:** Creating a dedicated "query refiner" agent that takes my vague request and outputs a precise, keyword-optimized search string before the `WebSurferAgent` executes it.

What's been your experience? Is the web search capability still a bit raw, or are there solid patterns to make it consistently useful for technical research?

--builder]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>backend_builder</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/troubleshooting-web-search-agent-returns-irrelevant-junk-how-to-improve-the-query-2/</guid>
                    </item>
				                    <item>
                        <title>Am I the only one who thinks the &#039;user proxy&#039; agent is the most useful part?</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/am-i-the-only-one-who-thinks-the-user-proxy-agent-is-the-most-useful-part-2/</link>
                        <pubDate>Tue, 25 Aug 2026 01:55:50 +0000</pubDate>
                        <description><![CDATA[Okay, hear me out. I&#039;ve been building workflows with AutoGen for a few months now, and I keep coming back to the same thought: the `UserProxyAgent` is the real MVP.

Everyone talks about the...]]></description>
                        <content:encoded><![CDATA[Okay, hear me out. I've been building workflows with AutoGen for a few months now, and I keep coming back to the same thought: the `UserProxyAgent` is the real MVP.

Everyone talks about the multi-agent conversations and the assistants, which are cool. But the user proxy? It's the bridge. It's what lets me, a human, actually *interact* with the whole system. Without it, I'm just watching bots talk to each other. With it, I can jump in, correct course, give approval, or just run that one-off code snippet. It turns a demo into a tool I can actually use. Anyone else feel like it's the unsung hero? &#x1f98a;]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>bluefox</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/am-i-the-only-one-who-thinks-the-user-proxy-agent-is-the-most-useful-part-2/</guid>
                    </item>
				                    <item>
                        <title>How do I integrate a simple REST API as a tool for an agent? The docs are confusing.</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/how-do-i-integrate-a-simple-rest-api-as-a-tool-for-an-agent-the-docs-are-confusing-2/</link>
                        <pubDate>Sun, 23 Aug 2026 10:20:57 +0000</pubDate>
                        <description><![CDATA[Trying to get an AutoGen agent to call a simple internal API. The official example is a bloated mess of decorators and nested classes for what should be a `curl` in a shell script.

Here&#039;s t...]]></description>
                        <content:encoded><![CDATA[Trying to get an AutoGen agent to call a simple internal API. The official example is a bloated mess of decorators and nested classes for what should be a `curl` in a shell script.

Here's the "simple" way I got it working after wasting an hour. Define your function, slap the `@tool` decorator on it. The agent's `llm_config` needs the tool spec registered.

```python
from autogen import AssistantAgent, UserProxyAgent, register_function
import requests

def get_weather(location: str) -&gt; str:
    """Call the weather API for a given location."""
    # Your actual API call here
    response = requests.get(f"http://internal-api/weather?city={location}")
    return response.text

# Create the agent that can use tools
assistant = AssistantAgent(
    name="assistant",
    llm_config={
        "tools": [{
            "type": "function",
            "function": {
                "name": "get_weather",
                "description": "Get current weather for a location",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "location": {"type": "string"}
                    },
                    "required": 
                }
            }
        }]
    }
)

user_proxy = UserProxyAgent(name="user_proxy", code_execution_config=False)

# This links the Python function to the tool spec
register_function(get_weather, caller=assistant, executor=user_proxy, name="get_weather")
```

Now the agent can call `get_weather`. It's just a wrapper. The docs make it seem like you need a PhD in their framework. You don't. It's a JSON schema and a function. The rest is theater.]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>devops_grunt</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/how-do-i-integrate-a-simple-rest-api-as-a-tool-for-an-agent-the-docs-are-confusing-2/</guid>
                    </item>
				                    <item>
                        <title>Best AutoGen setup for a hybrid on-prem and Azure deployment</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/best-autogen-setup-for-a-hybrid-on-prem-and-azure-deployment-2/</link>
                        <pubDate>Fri, 21 Aug 2026 13:45:56 +0000</pubDate>
                        <description><![CDATA[Hi everyone. We’re in the final stages of planning an enterprise deployment for AutoGen, and our environment is a bit of a split: some legacy services must remain on-prem (due to data sovere...]]></description>
                        <content:encoded><![CDATA[Hi everyone. We’re in the final stages of planning an enterprise deployment for AutoGen, and our environment is a bit of a split: some legacy services must remain on-prem (due to data sovereignty), while other components and the frontend will live in Azure.

I’m looking for advice on the most pragmatic architecture. Our core requirements are:
*   The ability for AutoGen agents to interface with both on-prem APIs/databases and Azure-hosted services (like Azure OpenAI).
*   Manageable networking and security without creating a nightmare of open ports.
*   A deployment model that doesn’t tie us to a single cloud vendor but acknowledges we're already Azure-heavy.

From my own war stories, the main pitfalls I foresee are latency between the two environments and the complexity of service discovery. I’ve seen teams get bogged down trying to make everything work over a VPN as if it were one flat network, and it never goes smoothly.

Has anyone successfully run a hybrid setup? I'm particularly interested in:
*   Where you placed the AutoGen "orchestrator" (the main runtime) – on a VM in Azure, or on-prem closer to the data?
*   How you handled authentication and secure communication between the segments.
*   Any specific Azure services (like Container Instances, App Service with VNet integration) that proved useful or problematic.

We're also in the middle of vendor/contract discussions for the Azure side, so any lessons on cost control for persistent agent workloads would be a bonus.

stay pragmatic]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>alexh42</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/best-autogen-setup-for-a-hybrid-on-prem-and-azure-deployment-2/</guid>
                    </item>
				                    <item>
                        <title>How do I set up a &#039;manager&#039; agent that actually evaluates other agents&#039; work effectively?</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/how-do-i-set-up-a-manager-agent-that-actually-evaluates-other-agents-work-effectively-2/</link>
                        <pubDate>Fri, 21 Aug 2026 02:50:53 +0000</pubDate>
                        <description><![CDATA[I&#039;m trying to build a review workflow where a manager agent checks the quality of responses from my support agents. Right now, my &quot;manager&quot; just approves everything. It doesn&#039;t seem to catch...]]></description>
                        <content:encoded><![CDATA[I'm trying to build a review workflow where a manager agent checks the quality of responses from my support agents. Right now, my "manager" just approves everything. It doesn't seem to catch errors or provide useful feedback.

What are the best practices for setting up a manager that can actually evaluate work? I'm using AutoGen for a SaaS helpdesk scenario. Should I be using specific evaluation functions, or is it more about the prompt design? Any concrete examples would be really helpful.]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>EmilyL</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/how-do-i-set-up-a-manager-agent-that-actually-evaluates-other-agents-work-effectively-2/</guid>
                    </item>
				                    <item>
                        <title>Troubleshooting: Agents sharing context incorrectly. Where is the state actually stored?</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/troubleshooting-agents-sharing-context-incorrectly-where-is-the-state-actually-stored-2/</link>
                        <pubDate>Wed, 19 Aug 2026 14:20:57 +0000</pubDate>
                        <description><![CDATA[Hi everyone! I&#039;m just starting out with AutoGen and ran into something confusing. &#x1f605;

I set up a simple two-agent chat (UserProxy and AssistantAgent), but it seems like they&#039;re sharin...]]></description>
                        <content:encoded><![CDATA[Hi everyone! I'm just starting out with AutoGen and ran into something confusing. &#x1f605;

I set up a simple two-agent chat (UserProxy and AssistantAgent), but it seems like they're sharing context in ways I didn't expect. When I ask the AssistantAgent about a file, it sometimes knows things from a previous conversation with the UserProxy that I thought were separate.

My main question is: **where is the conversational state actually stored?** Is it in memory, in a specific agent object, or somewhere else? I'm worried about accidentally leaking data between different task sessions.

Here's my basic setup:
```python
from autogen import AssistantAgent, UserProxyAgent

assistant = AssistantAgent("assistant", llm_config={...})
user_proxy = UserProxyAgent("user_proxy", code_execution_config={...})

user_proxy.initiate_chat(assistant, message="Explain Docker networking.")
# Later, in what I thought was a new context...
user_proxy.initiate_chat(assistant, message="What did we talk about earlier?")
# The assistant sometimes remembers the Docker conversation!
```

Could someone explain how this state management works in a beginner-friendly way? I'm coming from a Docker/Linux background where containers are isolated, so this is a new concept for me. Thanks in advance for any help!]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>devops_rookie_2025</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/troubleshooting-agents-sharing-context-incorrectly-where-is-the-state-actually-stored-2/</guid>
                    </item>
				                    <item>
                        <title>How do I version control and test my agent configurations? It feels like config spaghetti.</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/how-do-i-version-control-and-test-my-agent-configurations-it-feels-like-config-spaghetti-2/</link>
                        <pubDate>Tue, 18 Aug 2026 20:01:14 +0000</pubDate>
                        <description><![CDATA[Alright, I’m diving deep into AutoGen and I’ve hit a classic DevOps problem: **config sprawl**. My multi-agent setup is growing—orchestrators, specialists, tool-calling agents, each with the...]]></description>
                        <content:encoded><![CDATA[Alright, I’m diving deep into AutoGen and I’ve hit a classic DevOps problem: **config sprawl**. My multi-agent setup is growing—orchestrators, specialists, tool-calling agents, each with their own system prompts, temperature settings, function lists, and LLM configs. It’s starting to feel like the early days of unchecked Jenkinsfiles.

Right now, I’m just editing Python dictionaries and JSON blobs directly in my `agent_builder.py` or `config.py` files. This is… not scalable. I need to:
* Track changes to agent behaviors over time (did tweaking the planner’s prompt break our review workflow?).
* Test config changes in isolation before merging.
* Possibly have different configs for dev, staging, and prod (e.g., swapping GPT-4 for GPT-3.5 in dev for cost).
* Reuse and compose agent definitions across projects.

I’m thinking of a few approaches, but I’d love to hear what the community is doing.

**Option 1: Config-as-Code in Python**
Keep everything in Python files, but structure them as proper modules. Maybe:
```python
# agents/planner.yaml  (or .json, or .py as dict)
# Then load with something like:
import yaml
def load_agent_config(name, env="dev"):
    with open(f"agents/{name}.{env}.yaml") as f:
        return yaml.safe_load(f)
```
But then I need schema validation. Pydantic models for each agent type?

**Option 2: Versioned JSON/YAML in a dedicated config directory**
Treat it like Kubernetes manifests. Each agent is a manifest. Could use `kustomize`-like overlays for environment differences. This feels clean but requires a custom loader.

**Option 3: Everything in a dedicated class with inheritance**
BaseAgentConfig class, with specific agents overriding properties. Version control is just the code, but diffing becomes harder than with pure data files.

**My big questions:**
* How are you handling *secrets* within these configs (API keys aside, maybe personalization tokens)?
* Are you writing unit tests for your configs? Like, "ensure the summarizer agent never has code execution tools"?
* Has anyone built a pipeline that **lints** agent configs, or even does a dry-run of a conversation flow to validate before deploying?

I’m leaning towards YAML + Pydantic + a CI step that runs a validation script. Maybe even a snapshot test for key agent outputs to detect LLM provider drift. But I’m worried I’m over-engineering. Then again, isn’t that our thing? &#x1f605;]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>ci_cd_junkie</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/how-do-i-version-control-and-test-my-agent-configurations-it-feels-like-config-spaghetti-2/</guid>
                    </item>
				                    <item>
                        <title>My results after automating competitive pricing research. 80% accurate, needs human check.</title>
                        <link>https://communities.stackinsight.net/community/aitr-autogen/my-results-after-automating-competitive-pricing-research-80-accurate-needs-human-check/</link>
                        <pubDate>Tue, 18 Aug 2026 19:25:52 +0000</pubDate>
                        <description><![CDATA[Tried to use AutoGen to scrape and compare pricing pages for a few competitors. The agents could pull data, but the analysis was shaky.

It got the basic tiers right about 80% of the time. M...]]></description>
                        <content:encoded><![CDATA[Tried to use AutoGen to scrape and compare pricing pages for a few competitors. The agents could pull data, but the analysis was shaky.

It got the basic tiers right about 80% of the time. Missed the nuance every time. "Contact us" pricing was a total guess. It also couldn't parse enterprise deals with negotiated discounts, which is the whole game. Output looks clean but requires a full human review to be usable. Saves some manual copying, but that's it.]]></content:encoded>
						                            <category domain="https://communities.stackinsight.net/community/aitr-autogen/">AutoGen Reviews</category>                        <dc:creator>benjislack</dc:creator>
                        <guid isPermaLink="true">https://communities.stackinsight.net/community/aitr-autogen/my-results-after-automating-competitive-pricing-research-80-accurate-needs-human-check/</guid>
                    </item>
							        </channel>
        </rss>
		