Agent frameworks really were adopted. But the download curves also count dependency pulls, installer bursts and a moving floor, and three research groups now agree on where coordination pays: fan out the reading, keep one writer.
The README of Microsoft's AutoGen now opens with an orange badge that reads "maintenance mode". Underneath it is a plain sentence: "AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward." New users are sent to its successor, Microsoft Agent Framework.
AutoGen, from Microsoft Research, was one of the first frameworks built around agents talking to each other. Its own README says it "opened the door to experimental multi-agent orchestration patterns that inspired the community", and the community answered with stars. As of October 6, 2026 it has 61,269 of them on GitHub. That is more than CrewAI (59,381), more than LangGraph (42,757), and twice OpenAI's Agents SDK (29,852).
Now the other number. In July and August 2026, AutoGen's main package, autogen-agentchat, was downloaded from PyPI about a million times a month. LangGraph was downloaded 69 million times a month. Per star, that is about 16 downloads a month for AutoGen and about 1,600 for LangGraph, a hundredfold gap between how much a framework is admired and how much it gets installed.
"Everyone adopted multi-agent frameworks" is the story the charts tell, and the charts are real. The question worth asking is narrower: did the industry adopt these frameworks for the work where coordinating several agents actually helps? So I pulled three years of download data and set it beside the published evidence on when multi-agent systems win. The short answer is that the download curves cannot settle that question. What they can do is answer a sharper one, and it is the more useful of the two.
The data come from ClickHouse's public copy of the PyPI download logs. The obvious tool, pypistats, keeps only 180 days of history, which cannot draw a 2023-to-2026 curve. Raw PyPI counts also grow with Python itself: requests, the HTTP library that much of Python depends on, went from 877 million downloads in the first quarter of 2023 to 3.37 billion in the first quarter of 2026. So every number below is per 1,000 downloads of requests in the same period.
On that scale the adoption story is genuine:
openai-agents): from 0.07 when it appeared in early 2025 to 20.5.autogen-agentchat: it peaked at 0.93 in Q1 2026, the closest any quarter came to one per thousand, and was at 0.58 in July and August.But three things inside these curves are not anyone deciding to orchestrate agents.
The first is a dependency arrow. Every LangChain minor release from 1.0 to 1.4.3, the current release on October 6, 2026, lists LangGraph as a hard requirement (1.4.3 asks for langgraph>=1.2.11,<1.3.0), so a LangChain 1.x install pulls LangGraph with it. From August 1 to 24, LangChain's 1.x line was downloaded 73.0 million times and LangGraph's 1.x line 48.6 million. That is fewer for the dependency than for the library that requires it. Nothing in the count, then, shows LangGraph being chosen on its own. That the numbers cannot close at all is a reminder that downloads are not installs: caches, mirrors and lockfiles break the one-to-one.
The second is a burst. CrewAI was downloaded 28.4 million times in August 2026, about 16 times its September 2025 level. Of those, 21.3 million (75%) came through the uv installer. From August 14 to 23 it ran at 1.5 to 1.8 million downloads a day. On August 25 it was 91,000, and in September the month's total was 2.5 million, down 91%. An earlier spike, 14.5 million in May, came mostly through pip. The data do not say who or what was installing CrewAI a million and a half times a day, and I will not guess. What they do say is that a screenshot of that curve taken on August 24 would have been a very misleading picture of agent adoption.
The third is the floor moving. Between August and September 2026, packages that have nothing to do with agents fell almost as far:
requests fell 28.8%;numpy 27.8%;boto3 23.3%;openai client 31.0%;anthropic 24.9%.Daily requests downloads went from 66 to 68 million in the third week of August to between 27 and 50 million a day from August 25 through September. Any headline claiming that agent frameworks peaked in summer 2026 is reading a step in the whole index, whose cause is not in the data, as a verdict on agents.
There is also the thing no download can show. LangGraph is a graph runtime: you declare nodes and edges and it runs them. A graph with one model call in a loop and a graph with ten cooperating agents are the same download.
Put the two measures side by side and the frameworks rank almost in reverse:
| Framework | GitHub stars (Oct 6, 2026) | Average monthly downloads (Jul–Aug 2026) | Downloads per star per month |
|---|---|---|---|
| LangChain | 147,485 | 279.4M | about 1,900 |
| LangGraph | 42,757 | 69.2M | about 1,600 |
| OpenAI Agents SDK | 29,852 | 35.5M | about 1,200 |
| CrewAI | 59,381 | 20.4M (July alone: 12.4M) | about 340 (July: about 210) |
AutoGen (autogen-agentchat) | 61,269 | 1.0M | about 16 |
A star is a bookmark, given once, often while a project is new and exciting. AutoGen's repository dates from August 2023, when conversations between agents were the idea of the year. A download is a machine fetching the code, usually because something is about to import it. Per star, the two frameworks named for their multi-agent idea, crews and conversations, draw a fraction of what the general runtimes draw: CrewAI about a fifth of LangGraph's rate, even with a burst month in the average, and AutoGen about a hundredth. The steepest curves belong to libraries that run one agent as happily as ten.
Does that mean the industry turned away from multi-agent systems? Not by itself. It might mean people build multi-agent systems on general runtimes instead of on frameworks named for the pattern. To know which, you need the other half of the question: where coordination pays at all.
The answer has become unusually clear since mid-2025, because three independent groups measured it and arrived at the same split.
Anthropic described its multi-agent research system in June 2025.
Kim and colleagues turned the same intuition into a controlled study, "Towards a Science of Scaling Agent Systems". The current version is v3, April 2026.
Cognition, the company behind Devin, is the most interesting witness, because it changed its mind in public.
Put the three together and they agree on one rule: fan out the reading, keep one writer. Research is mostly reading. Code review is reading. Planning a sequence of dependent steps, editing one codebase, or holding one customer conversation is writing to a single shared state, and there the extra agents mostly add messages.
Computer architects have a name for the ceiling. In 1967 Gene Amdahl pointed out that the part of a job that must run serially caps the speedup from adding processors, however many you add. In an agent system the serial part is the synthesis: the one write, the final answer, the decision every subagent's work has to flow back into. Coordination messages do not shrink it. They add to it.
Two older fields found the same split in people.
Adam Smith opened The Wealth of Nations (1776) with a pin factory. The work was "divided into about eighteen distinct operations", and ten workers "could make among them upwards of forty-eight thousand pins in a day", where one alone might not make twenty. The famous part is the multiplication. The part people skip is the title of his third chapter: "That the Division of Labour is Limited by the Extent of the Market." Splitting work pays only when the work splits cleanly and there is enough of it to keep the specialists busy.
Fred Brooks supplied the cost side in The Mythical Man-Month (1975). A team in which every member must coordinate with every other has n(n−1)/2 lines of communication, so the talking grows with the square of the headcount while the hands grow only linearly. Hence his law: "Adding manpower to a late software project makes it later."
Swap workers for subagents and wages for tokens, and you have the recent research almost line for line:
None of this is new. What is new is that it can now be measured per task, with a benchmark and a token bill.
Here the answer has to be that the public data cannot say. No download records the task the framework runs, and no public dataset counts multi-agent deployments by the kind of work they do.
The closest public view comes from deployments, not downloads. LangChain's "State of Agent Engineering" survey drew 1,340 responses between November 18 and December 2, 2025:
Mapping those use cases onto the split is my inference, not the survey's. Customer service is one conversation with one customer, a single thread of state that every reply must stay consistent with: the tightly coupled case. Research and data analysis is the case Anthropic measured its 90.2% on: many sources read in parallel and one report written at the end. On that reading, the two biggest use cases sit on opposite sides of the line, in nearly equal shares.
What the download data do support is narrower, and it holds up:
That pattern is consistent with the evidence. The market appears to have kept the plumbing, runtimes on which you can build either a single agent or a lead agent with readers, and to have cooled on peer-to-peer agent conversation, the design Cognition now calls a distraction. Consistent with is not proof of. A team could be running a ten-agent swarm on LangGraph, and the download logs would never know.
That blind spot is the useful part. The industry's adoption chart cannot show whether multi-agent systems are being used where they help, and it was never going to: the thing that decides the answer, the serial fraction of the work, lives inside each task, not in a package index. A framework's download count tells you how often it was fetched. It does not tell you whether the work it runs could be split.
So the practical test for your own system is Amdahl's question, asked per task, before you pick a framework. How much of this job is reading that can fan out, and how much is the one write that cannot? Then make the comparison Anthropic's own data suggest: run your multi-agent design against a single agent given the same token budget. When token usage alone explains 80% of the variance, a multi-agent win that disappears at equal spend was a bigger budget, not a better architecture. If the win survives, you have found work that splits, and the evidence says to keep the writing single-threaded.
Sources
pypi.pypi_downloads_per_day, pypi.pypi_downloads_per_day_by_version, pypi.pypi_downloads_per_day_by_version_by_installer_by_type), sql-clickhouse.clickhouse.com, queried October 6, 2026 (data through October 5, 2026). Normalization per 1,000 requests downloads, and every figure in this essay, re-derived on that date.langchain 1.4.3 requires_dist: pypi.org/pypi/langchain/json, read October 6, 2026.The Agent Trust Stack gives each agent in a system a tamper-evident record of its actions, a portable reputation and a way to check the others' work, so a lead agent can tell which subagent's result to trust before it writes.
pip install agent-trust-stack npm install agent-trust-stack
For the provenance layer on its own: Hosted Chain of Consciousness.