AI agent framework release November 2026: browser-use 1.0.2 will not install next to Pydantic AI 2.67

browser-use 1.0.2 and Pydantic AI 2.67.0 will not install in the same environment. PyPI’s browser-use 1.0.2, uploaded at 2026-11-11T04:07:32Z, requires pydantic-ai-slim[anthropic,google,openai]==2.66.0. PyPI’s pydantic-ai 2.67.0, uploaded at 2026-11-11T00:23:27Z and read again at 04:12 UTC, requires pydantic-ai-slim[anthropic,cli,evals,google,logfire,mcp,openai,web]==2.67.0. Those are two exact pins on one library. A resolver cannot satisfy both.

That is the AI agent framework release that matters in this window. GitHub published browser-use 1.0.0 at 2026-11-10T02:04:48Z under the title “browser-use 1.0: CDP is all you need.” The agent no longer picks one action from a DOM dump. It writes Python in one persistent REPL on Chrome DevTools Protocol, on Pydantic AI. The 1.0.2 notes, published at 2026-11-11T04:00:00Z, are fixes on top of that rewrite. They do not loosen the pin.

How often these repositories tag a release, counted at 03:42 UTC the same morning, is 3,165 GitHub releases in 30 days, or 407 days of silence. That count does not include browser-use. The November 5 star table is AI agent framework GitHub stats. The bill for LangGraph, CrewAI, and the OpenAI Agents SDK, checked November 10, is LangGraph vs CrewAI vs the OpenAI Agents SDK.

Which packages actually published from November 9 to November 11?

The rows below are registry documents read between 04:07 and 04:12 UTC on November 11, 2026. A version outside that window stays in the table only so a pin does not get “upgraded” to a release that did not happen.

Source: aiagentsinfo.com. Times are upload_time_iso_8601 on PyPI or the npm time field, except google-genai’s Python row, which is the GitHub release published_at. HTTP Date on the browser-use document was Wed, 11 Nov 2026 04:08:36 GMT, when 1.0.2 was already latest.

Google ADK’s November 7 enum change, and the 30-day release counts, stay on the cadence note. This page is the behavior and the pin.

What does browser-use 1.0 change in the harness?

The loop. A harness that used to send a DOM snapshot and receive one click now hosts a Python cell. The 1.0.0 notes draw it as three steps: the model writes a cell, the REPL runs it on Chrome over CDP, the model calls finish(result). Tool use is that cell, not a fixed action schema. The library sits on Pydantic AI. The summary string on PyPI, read the same morning, is “Browser agent on Pydantic AI: one persistent Python REPL on raw CDP, cloud or local Chrome.”

The defaults move with the loop:

Source: aiagentsinfo.com, copied from the browser-use 1.0.0 release notes published November 10, 2026.

An option 1.0 cannot honor raises TypeError before a browser starts or anything is billed. Removed 0.x imports raise ImportError and name the replacement. except ImportError fallbacks still work. LangChain chat models raise TypeError. BROWSER_USE_ALLOWED_DOMAINS is not read: Browser() raises while it is set. After configuration checks pass, run() returns status='error' instead of raising, and a browser that started is stopped. A missing key still raises before anything is created.

Browser('chrome') attaches only to a browser whose live DevToolsActivePort is in your own profile directories. It no longer attaches to whatever answers on 127.0.0.1:9222. Tasks aimed at localhost or a private IP need Browser('chromium').

1.0.2, the PyPI latest that morning, fixes four things on top of 1.0.1. browser-use --cli-mcp drives its own Chromium when Chrome’s remote debugging is off, and tells the agent how to switch to the user’s Chrome. CDP calls from the CLI and MCP wait 20 seconds, matching the agent, and a page frozen by a JavaScript dialog fails with an error that says to handle the dialog. open() in MCP browser_exec reads and writes UTF-8 on Windows. browser-use feedback and a browser_feedback tool send scrubbed text.

GitHub’s license field on browser-use/browser-use was MIT at 04:11:03 GMT, with 120,632 stars. The download count through November 8, before 1.0, is Browser Use had 6,423,201 PyPI downloads. The directory listing is the product row.

Does the 33-point claim match the table on the same page?

The sentence and the table are both on the 1.0.0 notes. They are not the same comparison.

The sentence says: same models, up to 33 points higher, 2.7 to 5.2 times cheaper, 2 to 3 times faster. The table under it is BU Bench v2, 60 tasks times 3 reps, cloud browsers, dated 2026-11-09.

Source: aiagentsinfo.com, copied from the browser-use 1.0.0 release notes. Both arms ran gpt-6-luna on api.openai.com, judged by gpt-5.6-luna at xhigh. The default bu-fast model was not in this run.

An OpenAI outage hit the window. Pi lost 18 of 180 runs to provider errors. 1.0 lost 1. With those errors excluded, the paired difference of 1.0 minus Pi is score +1.1 [−3.2, +5.7], cost −$0.013 per task [−0.024, −0.004], and time −114 seconds [−202, −31]. The notes call the quality gap not significant and the cost and time gaps real. The 33-point sentence is still in the notes. It is not a row in this table. Treat the table as the number you can recompute a decision from, and the 33-point line as a claim whose comparison the page does not print.

result.cost_usd covers model and web-search charges. It does not cover the cloud browser or its proxy. The $0.070 is not a browser-hour bill.

The pricing page, read November 11, 2026, lists browsers at $0.02 per browser hour, a home-IP proxy at $5 per GB, your own proxy or a direct connection at $0.20 per GB, and the open-source harness at $0. A worked example on that page prices BU Fast as 20 million input tokens times $0.12 plus 1 million output tokens times $0.60, which is $3. bu-ultrafast is listed at $0.006 per task. Pay as you go is $0 a month, up to 1,000 browsers at once. Because Browser() is now the cloud browser, a 0.x script that called Browser() and ran on a local Chromium starts a meter it did not have.

A shopping agent that drives a page is this library plus a task, or a different product. Who can pay, on the paths already checked, is Can an AI shopping agent check out for you?. The dated product calendar is AI shopping agents, next 6 months.

Which pins does 1.0.2 refuse to move?

From requires_dist on the PyPI JSON for 1.0.2, read at 04:08:36 GMT:

Source: aiagentsinfo.com, from the PyPI JSON documents cited above.

The 1.0.0 notes say every dependency is pinned with ==, and that a project which needs a different pydantic, pydantic-ai-slim, mcp, google-genai, or pillow cannot install 1.0 beside it. They also say google-genai stays below 3 because 3.0.0 calls the Gemini API on v1, which rejects tools. The 1.0.2 document still shows google-genai==2.34.0, so that warning survived the patch. How to call gemini-4-carbon from a client that knows the id is Gemini 4 Carbon for agent harnesses. That call does not belong inside this pin.

What does Pydantic AI 2.67 change, once you are off the browser-use pin?

The GitHub tag v2.67.0 was published at 2026-11-10T23:11:23Z. The compatibility notes are the part a harness feels. This page did not re-run them.

RepoContext loads one instruction file per directory, and nested files, by default. walk_up is new. Coder builds on those files. Inlined nested instructions cap at 32,000 characters per conversation, with RepoContext(nested_max_chars=...). Nested instruction files surface when the model writes or edits in a directory. That is the file a coding agent treats as its standing orders. A cap of 32,000 characters is the point where a long AGENTS.md stops being free context.

SubAgents(dynamic_agents=...) works under durable execution. InputGuardrail, OutputGuardrail, and ToolGuardrail verdicts are recorded as durable operations. CapabilityCreation refuses to run when directory is below 'operator' authority. CodeMode rejects tools named after Python builtins, because the sandbox was calling the builtin instead of the tool. Provider-specific native-tool fields move to provider_settings. An OpenAI response that has no history_digest is no longer continued through openai_previous_response_id.

GitHub’s license field on pydantic/pydantic-ai was MIT at 04:11:03 GMT, with 21,131 stars. The library fee and the Logfire fee are separate. The directory listing keeps that split. The framework comparison is the wider table. A token bill for the model under the loop is the AI agent cost calculator.

If the environment also installs browser-use 1.0.2, stay on 2.66.0 until that pin moves. The 2.67 instruction-file behavior is not what that environment runs.

What did AG2 1.3.1 change on November 10?

The PyPI package ag2 1.3.1 uploaded at 2026-11-10T21:03:53Z. The GitHub tag is v.1.3.1, including the extra dot, published at 2026-11-10T21:03:11Z. The repository description calls AG2 formerly AutoGen. On the same morning, PyPI autogen was 0.14.1 from 2026-06-30, and pyautogen was 0.10.0 from 2025-07-15. Install ag2 if you want 1.3.1.

The notes add sandbox backends for SandboxShellTool and SandboxCodeTool: E2BEnvironment in ag2.extensions.e2b, and ag2.extensions.boxd for boxd Linux microVMs. MCP Apps work with browser-based hosts: /mcp no longer redirects POST or the CORS preflight. MCP-UI is documented as legacy. ag2.mcp_ui is not deprecated.

The upgrade note that will break a client: A2UI v1.0 now speaks the stable upstream spec. 1.3.0 targeted a release candidate. v1.0 messages carry "version": "1.0" instead of "v1.0". The basic catalog requires TextField.value. Custom v1.0 catalogs with a $ref the spec does not allow are rejected at load. protocol_version="v1.0" and the v0.9 default are unchanged.

Usage.cache_creation_extended_input_tokens counts OpenAI 12-hour and Anthropic 1-hour cache writes, which are priced differently, and the field flows through telemetry and eval traces. A context window bill that ignored extended cache writes was undercounting that line.

GitHub’s license field was Apache-2.0 at 04:11:45 GMT, with 5,056 stars. The homepage on that response was https://ag2.ai. Whether the old repository is maintained is Is AutoGen deprecated?. The directory listing is this package, not microsoft/autogen.

Which SDK bumps are real and still not a new agent loop?

OpenAI’s Python SDK 3.49.0, published at 2026-11-10T06:47:04Z, adds typed async function tools. That is a tool use typing change in the client. It is not a new Agents SDK. openai-agents was still 0.24.0, uploaded October 15. @openai/agents was still 0.21.0, published October 20.

Anthropic’s Python SDK 2.8.0, published at 2026-11-10T16:32:58Z, adds Claude Tag metrics and a reply_mode on Claude Tag Slack settings. The agent library is a different package. claude-agent-sdk 0.2.180 uploaded at 2026-11-09T20:08:02Z, and the GitHub release body is the install line. There is no behavior note on that tag to quote. The Claude Agent SDK listing is the product. How Fable 5.2 sits in that SDK is How to use Claude Fable 5.2 in an agent harness.

@langchain/langgraph 1.7.3 published at 2026-11-10T00:00:50Z. Python langgraph did not move. It was still 1.2.20 from November 5. GitHub’s latest release on langchain-ai/langgraph, read at 04:07 UTC, was still sdk==0.4.14 from November 6, a different package. Pin the package you import. The LangChain listing is the company. CrewAI 1.15.39, uploaded November 9, is the version already in the three-way bill. The CrewAI listing points there.

FastMCP 4.1.8, published at 2026-11-10T20:44:13Z, backports 28 fixes from the 4.2 line. The notes lead with OAuth proxy token revocation, which now works for the client that owns the token. They also fix recursive schema hydration, GitHub sessions across a transient verification failure, and proxied sampling. That is an MCP server bugfix, not a new spec. GitHub’s license field on PrefectHQ/fastmcp was Apache-2.0 at 04:11:03 GMT, with 28,377 stars. The directory listing is that repository.

google-genai 3.1.0 adds gemini-4-carbon to the Interactions model enum. browser-use 1.0.2 will not take that client. The twelve-month calendar of dated filings, which is not this week’s pin list, is AI agents, next 12 months.

What should you pin on November 11?

If you run browser-use, pin browser-use==1.0.2 and read MIGRATING.md before you bump. Keep pydantic-ai off 2.67.0 in that environment. Pass Browser('chromium') for localhost. Expect ImportError on 0.x imports, which is the point of the error: it names the replacement before a cloud browser starts.

If you run Pydantic AI and you do not install browser-use 1.0, 2.67.0 is the release with the instruction-file cap and the operator check. Read the compatibility list on the tag. The 32,000-character cap will truncate a nested instruction file you used to inline in full.

If you emit A2UI from AG2, change "v1.0" to "1.0" before you upgrade, and require TextField.value. If you run a FastMCP OAuth proxy, 4.1.8 is the revocation fix for the client that owns the token.

Leave openai-agents at 0.24.0 until that package actually publishes. Leave Python langgraph at 1.2.20 unless you meant the JavaScript package. The harnesses topic collects the longer guides. The best harness comparison is the product table, separate from these pins.

Frequently asked questions

What is the AI agent framework release in November 2026 that changes an install?

browser-use 1.0. GitHub published 1.0.0 at 2026-11-10T02:04:48Z. PyPI’s latest on the morning of November 11 was 1.0.2, uploaded at 2026-11-11T04:07:32Z. It requires pydantic-ai-slim==2.66.0. pydantic-ai 2.67.0, uploaded at 2026-11-11T00:23:27Z, requires pydantic-ai-slim==2.67.0. Those two exact pins do not resolve together.

Did browser-use 1.0 score 33 points higher?

The 1.0.0 release notes say the same models are up to 33 points higher, 2.7 to 5.2 times cheaper, and 2 to 3 times faster. The only score table on that page is BU Bench v2, dated 2026-11-09, comparing browser-use 1.0 in ultrafast mode with Browser Use Pi, both on gpt-6-luna. After provider errors are excluded, the paired score gap is +1.1 with a 95 percent interval from −3.2 to +5.7. That interval includes zero.

Which other agent frameworks published a release from November 9 to November 11, 2026?

PyPI ag2 1.3.1 uploaded at 2026-11-10T21:03:53Z. fastmcp 4.1.8 uploaded at 2026-11-10T20:44:54Z. crewai 1.15.39 uploaded at 2026-11-09T01:07:50Z. claude-agent-sdk 0.2.180 uploaded at 2026-11-09T20:08:02Z. The npm package @langchain/langgraph 1.7.3 was published at 2026-11-10T00:00:50Z. Python langgraph stayed at 1.2.20 from November 5. openai-agents stayed at 0.24.0 from October 15.

Does Browser() still launch a local browser in browser-use 1.0?

The 1.0.0 notes say Browser() starts a Browser Use Cloud browser. Browser('chromium') or Browser(local=True) selects a local Chromium. Browser('chrome') attaches only to a browser whose DevToolsActivePort file is in your own profile directories. The pricing page, read November 11, 2026, lists cloud browsers at $0.02 per browser hour.

What did AG2 1.3.1 change for agent builders?

The GitHub notes for tag v.1.3.1, published 2026-11-10T21:03:11Z, add E2B and boxd sandbox backends, fix MCP Apps on browser hosts, and change A2UI v1.0 so messages carry "version": "1.0" instead of "v1.0". The PyPI package to install is ag2, version 1.3.1. The packages named autogen and pyautogen were older distributions on the same morning.