<?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/">
  <channel>
    <title>DEV Community: Sarath Kavuru</title>
    <description>The latest articles on DEV Community by Sarath Kavuru (@sarathkavuru).</description>
    <link>https://dev.to/sarathkavuru</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4255618%2Fce18506b-c3ce-4fa8-aa0a-cd35958c0fed.png</url>
      <title>DEV Community: Sarath Kavuru</title>
      <link>https://dev.to/sarathkavuru</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/sarathkavuru"/>
    <language>en</language>
    <item>
      <title>AI Agent Framework Install Size: pip install agent-framework Pulls 912 MB, Its Core Package 18 MB (16 Packages Measured)</title>
      <dc:creator>Sarath Kavuru</dc:creator>
      <pubDate>Fri, 20 Nov 2026 13:53:23 +0000</pubDate>
      <link>https://dev.to/sarathkavuru/ai-agent-framework-install-size-pip-install-agent-framework-pulls-912-mb-its-core-package-18-mb-7e4</link>
      <guid>https://dev.to/sarathkavuru/ai-agent-framework-install-size-pip-install-agent-framework-pulls-912-mb-its-core-package-18-mb-7e4</guid>
      <description>&lt;p&gt;&lt;strong&gt;&lt;code&gt;pip install agent-framework&lt;/code&gt; put 207 Python distributions and 912.4 MB into a clean Python 3.13 environment on November 20, 2026. &lt;code&gt;pip install agent-framework-core&lt;/code&gt;, from the same Microsoft repository and the same 1.26.0 release, put in 11 distributions and 18.2 MB: a 50x difference for the same &lt;code&gt;from agent_framework import Agent&lt;/code&gt;.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That is the headline from a small experiment I ran this morning. I installed 16 packages from 14 Python AI agent frameworks, each into its own empty virtual environment, then measured what landed on disk and how long the main agent class took to import. The research question was simple: &lt;strong&gt;what does choosing an AI agent framework cost you in install size, dependency count, and cold-start import time before you write a line of agent code?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Install size rarely shows up in framework comparisons. It does show up in your container image, your Lambda or Cloud Run cold start, your CI cache, and your dependency audit.&lt;/p&gt;

&lt;h2&gt;
  
  
  How was the install footprint measured?
&lt;/h2&gt;

&lt;p&gt;Each package went into a fresh venv on one Linux box, and I measured that venv's &lt;code&gt;site-packages&lt;/code&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Sample:&lt;/strong&gt; 16 PyPI packages. That is the main package for LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, Microsoft Agent Framework, Pydantic AI, smolagents, LlamaIndex, Agno, Strands Agents, Haystack, AutoGen AgentChat, AG2 and DSPy, plus the slim variants &lt;code&gt;agent-framework-core&lt;/code&gt; and &lt;code&gt;pydantic-ai-slim&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Install:&lt;/strong&gt; &lt;code&gt;uv venv -p python3.13&lt;/code&gt; then &lt;code&gt;uv pip install &amp;lt;package&amp;gt;&lt;/code&gt; (uv 0.13.8), resolving the latest versions on PyPI on November 20, 2026, with no extras.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Size:&lt;/strong&gt; &lt;code&gt;du -sb&lt;/code&gt; of &lt;code&gt;site-packages&lt;/code&gt; after deleting every &lt;code&gt;__pycache__&lt;/code&gt;, so compiled bytecode is not counted. "Distributions" is the count from &lt;code&gt;uv pip list&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Import time:&lt;/strong&gt; wall-clock time of &lt;code&gt;python -c "&amp;lt;import the main agent class&amp;gt;"&lt;/code&gt;, run once to warm the bytecode, then 9 timed runs. I report the median. "Modules" is &lt;code&gt;len(sys.modules)&lt;/code&gt; after that import.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Machine:&lt;/strong&gt; x86_64 Linux, 8 vCPU, Python 3.13.5.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Limitations:&lt;/strong&gt; Wheels differ by platform, so macOS and Windows sizes will not match. Import times come from a shared cloud VM and will vary by a few tens of milliseconds between runs. Installing with pip instead of uv resolves the same dependency set, but pip adds itself to the venv. Extras such as &lt;code&gt;crewai[tools]&lt;/code&gt; were not installed, so real projects are often larger than this.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The table below is the full dataset: version, distribution count, size, largest top-level directory, import time and module count for every package.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which AI agent framework has the largest install?
&lt;/h2&gt;

&lt;p&gt;Microsoft Agent Framework's meta package is the largest at 912.4 MB, and CrewAI comes second at 604.7 MB.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Package (version)&lt;/th&gt;
&lt;th&gt;Distributions&lt;/th&gt;
&lt;th&gt;site-packages&lt;/th&gt;
&lt;th&gt;Largest directory&lt;/th&gt;
&lt;th&gt;Import median&lt;/th&gt;
&lt;th&gt;Modules loaded&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;agent-framework 1.26.0&lt;/td&gt;
&lt;td&gt;207&lt;/td&gt;
&lt;td&gt;912.4 MB&lt;/td&gt;
&lt;td&gt;claude_agent_sdk, 378.1 MB&lt;/td&gt;
&lt;td&gt;0.180 s&lt;/td&gt;
&lt;td&gt;322&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;crewai 1.15.44&lt;/td&gt;
&lt;td&gt;119&lt;/td&gt;
&lt;td&gt;604.7 MB&lt;/td&gt;
&lt;td&gt;lancedb, 189.9 MB&lt;/td&gt;
&lt;td&gt;1.731 s&lt;/td&gt;
&lt;td&gt;2,135&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;llama-index 0.14.26&lt;/td&gt;
&lt;td&gt;71&lt;/td&gt;
&lt;td&gt;184.3 MB&lt;/td&gt;
&lt;td&gt;numpy, 28.7 MB&lt;/td&gt;
&lt;td&gt;0.934 s&lt;/td&gt;
&lt;td&gt;1,331&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;pydantic-ai 2.71.0&lt;/td&gt;
&lt;td&gt;96&lt;/td&gt;
&lt;td&gt;127.9 MB&lt;/td&gt;
&lt;td&gt;cryptography, 14.9 MB&lt;/td&gt;
&lt;td&gt;0.735 s&lt;/td&gt;
&lt;td&gt;1,170&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;haystack-ai 3.5.0&lt;/td&gt;
&lt;td&gt;41&lt;/td&gt;
&lt;td&gt;97.4 MB&lt;/td&gt;
&lt;td&gt;numpy, 28.7 MB&lt;/td&gt;
&lt;td&gt;0.445 s&lt;/td&gt;
&lt;td&gt;1,067&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;strands-agents 1.65.0&lt;/td&gt;
&lt;td&gt;48&lt;/td&gt;
&lt;td&gt;64.5 MB&lt;/td&gt;
&lt;td&gt;botocore, 21.1 MB&lt;/td&gt;
&lt;td&gt;0.417 s&lt;/td&gt;
&lt;td&gt;917&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;google-adk 2.15.0&lt;/td&gt;
&lt;td&gt;48&lt;/td&gt;
&lt;td&gt;64.4 MB&lt;/td&gt;
&lt;td&gt;google, 24.7 MB&lt;/td&gt;
&lt;td&gt;0.686 s&lt;/td&gt;
&lt;td&gt;562&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;langgraph 1.2.24&lt;/td&gt;
&lt;td&gt;38&lt;/td&gt;
&lt;td&gt;53.3 MB&lt;/td&gt;
&lt;td&gt;zstandard, 23.9 MB&lt;/td&gt;
&lt;td&gt;0.560 s&lt;/td&gt;
&lt;td&gt;1,029&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;smolagents 1.26.0&lt;/td&gt;
&lt;td&gt;28&lt;/td&gt;
&lt;td&gt;52.5 MB&lt;/td&gt;
&lt;td&gt;pillow.libs, 14.7 MB&lt;/td&gt;
&lt;td&gt;0.327 s&lt;/td&gt;
&lt;td&gt;725&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;openai-agents 0.25.0&lt;/td&gt;
&lt;td&gt;38&lt;/td&gt;
&lt;td&gt;49.6 MB&lt;/td&gt;
&lt;td&gt;cryptography, 14.9 MB&lt;/td&gt;
&lt;td&gt;1.083 s&lt;/td&gt;
&lt;td&gt;1,358&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;agno 3.2.1&lt;/td&gt;
&lt;td&gt;31&lt;/td&gt;
&lt;td&gt;49.6 MB&lt;/td&gt;
&lt;td&gt;agno, 23.6 MB&lt;/td&gt;
&lt;td&gt;0.309 s&lt;/td&gt;
&lt;td&gt;480&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;autogen-agentchat 0.7.5&lt;/td&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;td&gt;30.4 MB&lt;/td&gt;
&lt;td&gt;pillow.libs, 14.7 MB&lt;/td&gt;
&lt;td&gt;0.231 s&lt;/td&gt;
&lt;td&gt;436&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;dspy 3.5.0&lt;/td&gt;
&lt;td&gt;26&lt;/td&gt;
&lt;td&gt;24.0 MB&lt;/td&gt;
&lt;td&gt;pydantic_core, 5.3 MB&lt;/td&gt;
&lt;td&gt;0.470 s&lt;/td&gt;
&lt;td&gt;796&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;pydantic-ai-slim 2.71.0&lt;/td&gt;
&lt;td&gt;17&lt;/td&gt;
&lt;td&gt;23.8 MB&lt;/td&gt;
&lt;td&gt;pydantic_ai, 10.9 MB&lt;/td&gt;
&lt;td&gt;0.499 s&lt;/td&gt;
&lt;td&gt;679&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;agent-framework-core 1.26.0&lt;/td&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;td&gt;18.2 MB&lt;/td&gt;
&lt;td&gt;pydantic_core, 5.3 MB&lt;/td&gt;
&lt;td&gt;0.160 s&lt;/td&gt;
&lt;td&gt;322&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ag2 1.3.3&lt;/td&gt;
&lt;td&gt;14&lt;/td&gt;
&lt;td&gt;16.6 MB&lt;/td&gt;
&lt;td&gt;pydantic_core, 5.3 MB&lt;/td&gt;
&lt;td&gt;0.242 s&lt;/td&gt;
&lt;td&gt;531&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Measured November 20, 2026. Python 3.13.5, uv 0.13.8, Linux x86_64. Sizes exclude &lt;code&gt;__pycache__&lt;/code&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;The spread runs from 16.6 MB (AG2) to 912.4 MB, a 55x range across packages that all give you an agent loop with tool calling.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why is &lt;code&gt;pip install agent-framework&lt;/code&gt; 912 MB?
&lt;/h2&gt;

&lt;p&gt;The meta package installs every integration, and one of them bundles a full Claude Code binary.&lt;/p&gt;

&lt;p&gt;The &lt;code&gt;agent-framework&lt;/code&gt; 1.26.0 package declares a single requirement, &lt;code&gt;agent-framework-core[all]==1.26.0&lt;/code&gt;. The &lt;code&gt;all&lt;/code&gt; extra pulls in 29 &lt;code&gt;agent-framework-*&lt;/code&gt; integration packages: connectors for Anthropic, Bedrock, Gemini, Mistral, Ollama, Redis, Cosmos DB, Foundry, Copilot Studio, Purview, DevUI and more. One of them, the Claude connector, depends on &lt;code&gt;claude-agent-sdk&lt;/code&gt; (0.2.186 in this install). That package's &lt;code&gt;_bundled/&lt;/code&gt; folder holds a &lt;code&gt;claude&lt;/code&gt; executable whose version file reads &lt;code&gt;2.1.329 (Claude Code)&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;That one directory is 378.1 MB, or &lt;strong&gt;41% of the whole environment&lt;/strong&gt;. After it, the next largest directories are &lt;code&gt;onnxruntime&lt;/code&gt;, &lt;code&gt;python_guest&lt;/code&gt;, &lt;code&gt;foundry_local_sdk&lt;/code&gt;, &lt;code&gt;numpy&lt;/code&gt; and &lt;code&gt;botocore&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;If you only need one or two model providers, the practical fix is to install &lt;code&gt;agent-framework-core&lt;/code&gt; plus the specific &lt;code&gt;agent-framework-&amp;lt;provider&amp;gt;&lt;/code&gt; packages you call. The import path is the same: &lt;code&gt;from agent_framework import Agent&lt;/code&gt; worked in both environments. Because the meta package loads its integrations lazily, its import time (0.180 s) is almost identical to core's (0.160 s). Neither one loaded more than 322 modules to get &lt;code&gt;Agent&lt;/code&gt;. The cost of the meta package is disk, image size and supply-chain surface, not startup time.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why does CrewAI take 1.7 seconds to import?
&lt;/h2&gt;

&lt;p&gt;CrewAI eagerly loads 2,135 modules, which is more than any other package in this test, and it requires a vector-database stack.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;crewai&lt;/code&gt; 1.15.44 lists &lt;code&gt;lancedb&lt;/code&gt; and &lt;code&gt;qdrant-client&lt;/code&gt; as hard requirements, not extras. LanceDB requires &lt;code&gt;pyarrow&lt;/code&gt;. Together, &lt;code&gt;lancedb&lt;/code&gt; (189.9 MB) and &lt;code&gt;pyarrow&lt;/code&gt; (171.2 MB) are 361.1 MB of CrewAI's 604.7 MB, or 60%. LanceDB and Qdrant are both vector stores, the kind of dependency that backs agent memory and knowledge retrieval. Every CrewAI install pays for them, whether or not your crew uses those features.&lt;/p&gt;

&lt;p&gt;The import number matters more than the disk number for short-lived workers. A 1.731 s median to run &lt;code&gt;from crewai import Agent, Crew&lt;/code&gt; is 1.7 s added to every cold start in a serverless function or a per-task container. The next slowest were the OpenAI Agents SDK (1.083 s) and LlamaIndex (0.934 s).&lt;/p&gt;

&lt;h2&gt;
  
  
  Does a bigger install mean a slower import?
&lt;/h2&gt;

&lt;p&gt;No. In this sample, disk size and import time barely track each other.&lt;/p&gt;

&lt;p&gt;The largest install (agent-framework, 912.4 MB) imported in 0.180 s. The OpenAI Agents SDK, the 10th-largest install at 49.6 MB, took 1.083 s and loaded 1,358 modules. Lazy imports decide import time, and the dependency list decides disk size. You have to measure both, because neither predicts the other.&lt;/p&gt;

&lt;p&gt;GitHub popularity does not predict either one. In aiagentsinfo.com's &lt;a href="https://www.aiagentsinfo.com/reports/ai-agent-framework-github-stats-november-2026" rel="noopener noreferrer"&gt;GitHub stats snapshot of 19 agent framework repositories&lt;/a&gt;, taken on November 5, 2026, crewAIInc/crewAI had 60,514 stars and pydantic/pydantic-ai had 21,032. On install weight their order flips: CrewAI is 604.7 MB, and Pydantic AI's full package is 127.9 MB, or 23.8 MB as &lt;code&gt;pydantic-ai-slim&lt;/code&gt;. That same snapshot found that 14 of the 19 repositories had a release within 30 days. That is one more reason to re-measure after each upgrade rather than trust last quarter's image size.&lt;/p&gt;

&lt;h2&gt;
  
  
  Are the slim packages worth it?
&lt;/h2&gt;

&lt;p&gt;Yes, if you know which providers you need. The slim packages cut size by 5x to 50x without changing your import line.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;pydantic-ai&lt;/code&gt; â†’ &lt;code&gt;pydantic-ai-slim&lt;/code&gt;: 96 â†’ 17 distributions, 127.9 â†’ 23.8 MB (5.4x smaller), import 0.735 â†’ 0.499 s. The full package pulls in the OpenAI, Anthropic and Google GenAI SDKs, OpenTelemetry, &lt;code&gt;tiktoken&lt;/code&gt; and MCP. The slim one lets you add only the extras you use, for example &lt;code&gt;pydantic-ai-slim[openai]&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;agent-framework&lt;/code&gt; â†’ &lt;code&gt;agent-framework-core&lt;/code&gt;: 207 â†’ 11 distributions, 912.4 â†’ 18.2 MB (50x smaller).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Be careful with download counts here. aiagentsinfo.com's &lt;a href="https://www.aiagentsinfo.com/trackers/downloads/pydantic-ai-slim" rel="noopener noreferrer"&gt;daily PyPI download tracker for pydantic-ai-slim&lt;/a&gt; recorded 9,010,191 downloads from November 13 to 19, 2026, against 1,577,815 for &lt;code&gt;pydantic-ai&lt;/code&gt;. But &lt;code&gt;pydantic-ai&lt;/code&gt; depends on the slim package, so every full install also downloads the slim one. That gap counts dependency installs, CI runs and direct slim users together. It does not show how many teams chose slim on purpose. The same caveat applies to the &lt;a href="https://www.aiagentsinfo.com/trackers/downloads/crewai" rel="noopener noreferrer"&gt;CrewAI download history&lt;/a&gt; and the rest of the &lt;a href="https://www.aiagentsinfo.com/trackers/downloads" rel="noopener noreferrer"&gt;framework downloads tracker&lt;/a&gt;: a download is a file request, not a project.&lt;/p&gt;

&lt;h2&gt;
  
  
  How should this change which framework you pick?
&lt;/h2&gt;

&lt;p&gt;Install footprint should break ties, not make the decision. The capability you need comes first.&lt;/p&gt;

&lt;p&gt;A graph runtime with checkpointing, a role-based crew, and a minimal handoff loop solve different problems. The &lt;a href="https://www.aiagentsinfo.com/blog/langgraph-vs-crewai-vs-openai-agents-sdk" rel="noopener noreferrer"&gt;comparison of LangGraph, CrewAI and the OpenAI Agents SDK&lt;/a&gt; and the broader &lt;a href="https://www.aiagentsinfo.com/best/ai-agent-frameworks" rel="noopener noreferrer"&gt;guide to the best AI agent frameworks&lt;/a&gt; cover that side. If protocol support decides it for you, there is also a &lt;a href="https://www.aiagentsinfo.com/research/ai-agent-framework-mcp-a2a-support-november-2026" rel="noopener noreferrer"&gt;check of which frameworks support MCP and A2A&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Once you have two or three candidates, the numbers above turn into a short checklist:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Container or serverless?&lt;/strong&gt; Avoid meta packages. Install the core package plus the provider packages you use, and run &lt;code&gt;du -sh $(python -c "import site; print(site.getsitepackages()[0])")&lt;/code&gt; in CI so an upgrade that adds 300 MB fails a size budget instead of shipping silently.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cold starts matter?&lt;/strong&gt; Time your real import line with &lt;code&gt;python -X importtime -c "..."&lt;/code&gt;. In this run, 1.7 s (CrewAI) and 0.16 s (agent-framework-core) both count as "fast" in a notebook, but they behave very differently in a function that starts 10,000 times a day.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Security review?&lt;/strong&gt; The distribution count is your dependency audit surface. 207 packages and 11 packages are very different reviews, even when they come from the same vendor.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check for bundled binaries.&lt;/strong&gt; A 378 MB executable inside a Python wheel is easy to miss in &lt;code&gt;pip list&lt;/code&gt; and obvious in &lt;code&gt;du&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What's the takeaway?
&lt;/h2&gt;

&lt;p&gt;The same agent class can cost you 18 MB or 912 MB, depending on which package name you type. In both of the full-versus-slim pairs I measured, the obvious package name was the heavy one. Measure your own install before you ship.&lt;/p&gt;

&lt;p&gt;I'll re-run this measurement as new versions land. If you maintain one of these frameworks and think a number is wrong, the install and import commands are listed above so you can reproduce it.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;This article was written with AI assistance and reviewed for accuracy.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>python</category>
      <category>ai</category>
      <category>agents</category>
      <category>performance</category>
    </item>
  </channel>
</rss>
