The open-source agent frameworks consolidation — 2026 edition.

Who stayed: LangChain, LangGraph, CrewAI, Continue

LangChain entered 2026 having absorbed the most sustained criticism of any framework in the segment — a recurring complaint about the abstraction depth, the breaking changes between minor versions, and the cost of maintaining LangChain dependencies as the framework evolved. The criticism has been substantively addressed by Harrison Chase and the LangChain team through three structural changes that the PyPI dependency data captures. The first is the stabilisation of LangChain Core, the minimal-dependency package that LangChain v0.3 split out as a stable foundation on which the broader ecosystem could build. The second is the bifurcation of the integration surface into the langchain-community package, which absorbed the volatility of provider-specific integrations and allowed the core package to maintain a more stable API surface. The third is the migration of the orchestration logic into LangGraph, the project Chase and the team launched in early 2024 and which has become the layer that the more sophisticated production deployments actually depend on.

The LangGraph download trajectory is the clearest evidence of the segment's commercial maturation. PyPI weekly downloads on LangGraph rose from approximately 180,000 in May 2025 to approximately 740,000 in May 2026 — a 4.1× increase against a baseline that already represented meaningful share. The npm package @langchain/langgraph, which carries the JavaScript and TypeScript bindings, grew from 67,000 weekly downloads to 290,000 over the same window. The growth is concentrated in production deployments rather than experimental usage; the LangSmith trace data that LangChain Inc. publishes in its quarterly ecosystem update shows that LangGraph is now the orchestration layer underneath a majority of the agent workflows running in production at LangSmith's enterprise customers. The structural position is unambiguous: LangGraph has become the orchestration default for the segment.

CrewAI's trajectory tells a narrower but similar story. João Moura's framework, which positioned itself early as the multi-agent coordination layer, has held its share in the multi-agent segment despite increasing competition from LangGraph's multi-agent primitives and from custom multi-agent architectures that enterprise platform teams have built directly. PyPI weekly downloads on the crewai package rose from approximately 95,000 in May 2025 to approximately 155,000 in May 2026 — a slower growth rate than LangGraph but a meaningful absolute gain. The growth is concentrated in two segments: smaller enterprise teams that prefer CrewAI's more opinionated abstraction over LangGraph's lower-level primitives, and a substantial education-and-tutorial segment that has adopted CrewAI as the reference framework for teaching multi-agent concepts. The strategic posture of CrewAI Inc. — the commercial entity Moura raised a Series A around in mid-2025 — has stabilised the project's commit velocity and produced the kind of governance maturity that enterprise buyers require. The PyPI data and the GitHub commit data both reflect this.

Continue, the IDE-integration project that Ty Dunn and Nate Sesti maintain, has absorbed the segment that the broader framework projects ceded as their attention moved toward production runtime concerns. Continue's npm weekly downloads grew from approximately 45,000 to approximately 180,000 over the year, and the VS Code marketplace installs crossed 1.2 million in April 2026. The project's strategic positioning — open-source IDE integration with a commercial layer for team-scale deployments — has produced a sustainable revenue trajectory that the founding team has used to fund continuing development without compromising the open-source license. Continue's structural position in the segment is not that it competes with Cursor or Windsurf on the closed-source IDE-integration market, but that it provides the open-source alternative for organisations that require source-available tooling or that have specific security postures that exclude closed-source agents. The segment is smaller than the closed-source market but is well-defined and has produced a durable commercial position for the Continue team.

Who folded: OpenInterpreter, Open Devin, Plandex

OpenInterpreter, the project Killian Lucas founded in mid-2023 and that briefly attracted some of the most concentrated developer attention in the segment, has lost commit velocity at a rate that the PyPI data captures unambiguously. The project's weekly downloads peaked at approximately 78,000 in early 2024 and have declined to approximately 12,000 in May 2026. The GitHub commit cadence has fallen from a median of 23 commits per week in 2024 to a median of 4 commits per week through Q1 2026, and the contributor count on the main branch has narrowed from a peak of 47 active contributors to 8. The project's positioning — a local-execution agent that runs a Python interpreter inside the agent loop — was attention-catching in the early agent moment but has been substantially absorbed by the broader Claude Code, Cursor Composer, and Aider tooling that offers a similar capability surface with more mature integration. Lucas has been candid in public posts that his attention has moved toward a different product direction, and the OpenInterpreter project has effectively entered maintenance mode without a formal sunset announcement.

Open Devin, the project that emerged from the All Hands AI team in early 2024 as an open-source response to Cognition Labs' Devin product, has had a more complicated trajectory. The project rebranded to OpenHands in late 2024 and the All Hands team has continued to invest in it, but the strategic positioning has narrowed from the original "open-source autonomous software engineer" framing to a more specific positioning as a coding agent platform that competes more directly with Aider and Cursor than with the broader autonomous-engineer category. The download data reflects the rebranding and the strategic shift: the original opendevin package on PyPI has effectively zero weekly downloads in May 2026 (developers migrated to the new package); the openhands package shows approximately 38,000 weekly downloads, well below the comparable LangGraph figure but above OpenInterpreter's residual usage. The project remains active and well-maintained, but its share of the segment is narrower than the original Open Devin positioning suggested. The All Hands team's strategic posture — Series A from major venture investors, focused commercial product around code agents — provides the resources for continued development, but the segment competition is more intense than the original entry positioning anticipated.

Plandex, the project Daniel Tate launched in mid-2024 as a CLI-based agent for complex software engineering tasks, has lost its early momentum more sharply than the project's public communications would suggest. The PyPI download data shows weekly downloads peaking at approximately 18,000 in late 2024 and declining to approximately 3,200 in May 2026. The commit cadence has fallen substantially, and the contributor count has narrowed to a core team of two. Tate's attention has shifted toward a commercial product offering that uses Plandex internally rather than as a standalone open-source reference, and the open-source distribution has effectively become a marketing channel for the commercial product rather than a primary product surface. The structural position is that Plandex is no longer a primary reference in the segment, though the underlying technology continues to be used in Tate's commercial offering. The implication for developers evaluating frameworks is that Plandex should be considered a project whose open-source release cadence will continue but whose contribution surface has narrowed to a level that production deployments would be advised to weight in their evaluation.

The pattern across the three projects is informative. None of them produced a structural failure in their technology approach; the technical positions they staked out have been substantially validated by subsequent product evolution in the segment. What they did not produce was a commercial backing entity strong enough to sustain the development cadence the segment now requires. The competition for developer attention has compressed substantially as enterprise buyers have moved toward production deployments that require the kind of governance, observability, and integration maturity that takes a funded team to deliver. Projects that depend on founder-team volunteer cycles cannot sustain the cadence; projects that have a commercial entity behind them can. The structural finding is that the segment is no longer in the founder-driven open-source phase. It has entered the commercially-backed open-source phase, and the projects that have not made that transition are losing share.

Who acquihired: AutoGen, Goose, and the Microsoft pattern

AutoGen, the framework that Microsoft Research released in late 2023 and that the research community treated as a primary reference through 2024, has had a complicated 2025 and 2026. The original Microsoft Research team — including Chi Wang and Qingyun Wu, the core researchers who built the framework — left Microsoft in mid-2024 to start AG2, a commercial entity built around the AutoGen architecture. Microsoft continued to maintain the AutoGen repository under its own governance, and the resulting fork created confusion in the developer community about which project represented the canonical AutoGen direction. The download data reflects the confusion: weekly downloads on the autogen PyPI package have remained relatively stable at approximately 65,000, but the package now carries both lineage histories and the version trajectories of the Microsoft-maintained and AG2-maintained variants have diverged in ways that production users have had to track carefully.

The acquihire pattern that the segment has produced is the more interesting structural development. Block Inc., the company that absorbed the engineering team behind Goose — the agent framework that emerged from Block's developer experience team in 2024 — has continued to invest in Goose as an open-source project while integrating the underlying capabilities into Block's internal engineering platform. The Goose npm download trajectory has been less consequential than the broader framework competition, but the project's commercial backing has held the commit cadence and the contributor count at a level that the more attention-driven projects have not maintained. The Goose model — an open-source project with a clear commercial entity providing engineering resources without converting the project to a closed-source product — has become a template that several other framework projects have studied as a strategic option.

The Microsoft pattern around AutoGen — original research team leaves to start a commercial entity, original employer continues to maintain the repository, the fork creates ecosystem confusion that takes 12-18 months to resolve — is the cautionary tale the segment has internalised. The AG2 project, which Chi Wang and Qingyun Wu have positioned as the canonical continuation of the AutoGen architecture, has produced a substantively stronger commit cadence than the Microsoft-maintained variant through 2025 and 2026, but the dual-lineage problem has cost both variants share to LangGraph and CrewAI in the production-deployment segment. The structural lesson the segment has drawn is that founder-team continuity matters more than license or repository ownership for the medium-term trajectory of an open-source framework, and that the financial decisions of the original employer can produce ecosystem-level fragmentation that the technical merits of either variant cannot recover from quickly.

The Cline and RooCode projects represent a different pattern that has become consequential in the IDE-integration segment. Cline, the project that Saoud Rizwan maintains, has grown from approximately 12,000 weekly npm downloads in early 2025 to approximately 95,000 in May 2026, with the VS Code marketplace install base crossing 800,000 in April. RooCode, which forked from Cline in early 2025 under the leadership of a separate maintainer team, has grown to approximately 38,000 weekly downloads. The two projects compete in a segment that the original framework projects do not directly address — the IDE-integrated agent assistant — and the competition between them has been productive in the sense that both projects' feature velocity has been higher than would have been the case for either project alone. The structural position is that the IDE-integration segment has been ceded by the broader framework projects (LangChain, LangGraph, CrewAI) and absorbed by a set of projects optimised specifically for that integration surface. Continue, Cline, and RooCode together carry the open-source IDE-integration segment.

The dependency graph shifts that the download data understates

The PyPI and npm download data captures the headline volume but understates the structural importance of the dependency graph shifts that the framework consolidation has produced. The most consequential shift is the deepening of LangGraph as a transitive dependency of the broader Python agent ecosystem. The dependency analysis the LangChain team published in its April 2026 ecosystem report shows that approximately 23 per cent of the production agent applications running on PyPI dependencies as of Q1 2026 have LangGraph as a direct or transitive dependency. The figure was 8 per cent in Q1 2025. The growth reflects both LangGraph's own download growth and the migration of secondary frameworks — Haystack, LlamaIndex, and several smaller projects — to use LangGraph as their orchestration substrate rather than maintaining proprietary orchestration logic. The transitive dependency footprint is the structural moat that the LangChain Inc. commercial entity has built around the LangGraph project, and it is the moat that the alternative orchestration projects have not been able to match.

The npm dependency data tells a similar story with different actors. The @langchain/langgraph package's dependency footprint in the JavaScript ecosystem has grown from approximately 4 per cent of production agent applications to 17 per cent over the same window. The growth is concentrated in the Next.js and Remix application segment where Vercel's AI SDK has driven much of the developer momentum. The structural position is more contested in the npm ecosystem than in the Python ecosystem, because the Vercel AI SDK provides an orchestration alternative that some developers prefer for its tighter integration with the Vercel deployment platform. The competition between LangGraph and the Vercel AI SDK at the orchestration layer is the most consequential ongoing competition in the npm side of the segment, and the outcome will likely be shaped by which platform — LangChain's commercial offering or Vercel's deployment platform — produces stronger enterprise-level integration over the next twelve months.

The dependency graph shifts on the abandoned projects are equally informative. OpenInterpreter's transitive dependency footprint, which peaked at approximately 1.2 per cent of production Python agent applications in mid-2024, has fallen below 0.1 per cent in May 2026. The projects that depended on OpenInterpreter as a runtime have substantially migrated to alternative architectures, and the projects that depended on OpenInterpreter for development tooling have migrated to Aider or Continue or to the closed-source alternatives. The dependency graph atrophy is the leading indicator of project abandonment that the download data lags by approximately one quarter, because production deployments are conservative about migration but eventually do migrate when the project's commit cadence drops below a threshold the platform team can defend.

Plandex's dependency graph contraction has been sharper because the project's commercial-backing transition pulled the open-source distribution into a more constrained role. The plandex package's transitive dependency footprint has fallen to approximately 0.02 per cent of production Python agent applications, a level at which the project effectively functions as a niche reference rather than a substrate. The Open Devin / OpenHands trajectory is intermediate: the dependency footprint has stabilised at approximately 0.4 per cent of production applications, well below the LangGraph or CrewAI level but well above OpenInterpreter's residual usage. The structural position is that OpenHands is a primary reference for the specific segment it now targets but is not a general-purpose framework in the way the original Open Devin positioning suggested. The dependency graph data confirms the narrative the All Hands team has communicated; the headline download number understates the strategic shift.

The strategic posture of each project's backing entity

The structural finding of the consolidation analysis is that the projects that have survived and grown share are uniformly the projects whose backing entity has a coherent commercial strategy and a sustainable funding position. LangChain Inc., the commercial entity behind LangChain and LangGraph, raised a Series A and Series B from major venture investors and has built its commercial product — LangSmith for observability, the LangChain Platform for managed deployment — around a strategy that does not require closing the open-source projects. The strategy is sustainable in the sense that LangSmith's enterprise revenue funds the engineering investment in the open-source frameworks, and the open-source frameworks compound the LangSmith funnel. The economic flywheel has produced the most coherent open-source strategy in the segment, and the segment's consolidation toward LangChain has been the structural consequence.

CrewAI Inc., the commercial entity João Moura built around the CrewAI framework, has produced a similar but smaller-scale flywheel. The commercial product — CrewAI Enterprise, which provides hosted multi-agent orchestration with audit logging and team management — funds the engineering investment in the open-source framework. The strategy is sustainable in the multi-agent segment but the segment's growth ceiling is below the orchestration ceiling LangGraph operates in, and CrewAI Inc.'s long-term position depends on either growing the multi-agent segment or expanding into adjacent categories that the framework's positioning does not currently address. Moura has been candid in public presentations that the company's roadmap includes adjacent product offerings; the open-source community will watch the adjacent expansion closely because the most common failure mode for commercial-backed open-source projects is the dilution of the open-source community as the commercial entity prioritises adjacent products over the core framework.

All Hands AI, the entity behind OpenHands (formerly Open Devin), has positioned itself in the more contested segment of coding-agent platforms where the commercial alternatives — Cursor, Cognition Devin, Claude Code, the broader closed-source segment — are well-funded and growing share. The All Hands team's strategy depends on producing an open-source product that enterprise customers will adopt for specific reasons closed-source alternatives do not satisfy: source availability for security or regulatory reasons, on-premise deployment requirements, or specific customisation needs that the closed-source products do not address. The strategy is plausible but narrow, and the All Hands team's commercial trajectory through the next twelve months will be the data point that resolves whether OpenHands holds its present share or contracts further. The team's Series A funding provides runway for the strategic experiment to play out, but the segment's competitive intensity is the principal risk.

Block Inc.'s strategy around Goose is the model the segment will likely converge toward. Block treats Goose as an open-source contribution that compounds Block's broader engineering platform investment, without depending on Goose as a primary revenue driver. The strategy decouples the open-source project's funding from its direct revenue contribution, which produces a more stable engineering investment than the project-as-revenue-driver model the smaller commercial entities depend on. The pattern is reproducible for any large engineering organisation with an internal platform investment that produces tooling worth releasing publicly. The pattern has produced the most stable open-source contribution among the projects that started in 2023 and 2024, and the structural lesson for the segment is that founder-team or research-team led projects are less stable than corporate-platform-team led projects because the funding model is more contingent on the team's personal attention budget. The Continue project's positioning is closer to the Goose model than to the LangChain or CrewAI models, and Continue's trajectory through 2026 will test whether the founder-team can sustain a Goose-like stability without the corporate platform investment that backs Block's contribution.

What to watch

The segment will continue to consolidate through 2026 and into 2027, and the dependency graph data will lead the download data in signalling which projects are losing share before the download data confirms it. The October 2026 ecosystem reports from LangChain Inc. and CrewAI Inc. will be the next significant data points, and the trajectory of the projects that have not produced a commercial backing entity will be the most diagnostic signal of the segment's continued direction.

• Whether the competition between LangGraph and the Vercel AI SDK at the npm orchestration layer produces a clear winner over the next twelve months, or whether the segment splits along the deployment-platform axis with LangGraph dominating non-Vercel deployments and the Vercel AI SDK dominating Vercel deployments.
• Whether the AutoGen / AG2 fork resolves toward a canonical project — either through Microsoft sunsetting its variant or through AG2 establishing dominant share — or whether the dual-lineage problem continues to cost both variants share to LangGraph.
• Whether the IDE-integration segment that Continue, Cline, and RooCode contest produces a consolidation toward one project or maintains the three-way competition; the segment's dynamics resemble the early-stage framework competition from 2023 and the same consolidation pressures are likely to apply.
• Whether new entrants emerge in the segment with the funding and commercial backing required to compete with the present incumbents; the segment's entry barrier has risen substantially as enterprise buyers have moved toward production deployments, and new projects without commercial backing are unlikely to produce a meaningful trajectory.
• Whether the LangChain Inc. commercial strategy holds up under the competitive pressure of Anthropic's deeper integration of agent capabilities into the Claude product surface; LangChain Inc.'s position depends on LangSmith remaining a category-defining observability product, and Anthropic's product trajectory has the most potential to compress that category.

Frequently asked

Why is LangGraph treated as more structurally important than LangChain itself in this analysis?

The LangChain v0.3 split moved the orchestration logic — the substantive runtime capability that production deployments depend on — out of the langchain package and into the LangGraph package. LangChain remains the lower-level abstraction and integration layer, but the projects building production agent applications are predominantly depending on LangGraph for orchestration. The PyPI download trajectory reflects this: LangGraph's growth has outpaced LangChain's, and the transitive dependency analysis shows LangGraph deeper in the production application stack than the broader LangChain package. The structural position is that LangGraph has become the orchestration default; LangChain remains the broader ecosystem, but the load-bearing element is LangGraph.

What is the AutoGen / AG2 fork, and why has it not resolved?

The original Microsoft Research team behind AutoGen — including Chi Wang and Qingyun Wu — left Microsoft in mid-2024 to start AG2, a commercial entity built around the AutoGen architecture. Microsoft continued to maintain the AutoGen repository under its own governance. The two variants have diverged in their version trajectories and feature priorities. The fork has not resolved because neither variant has produced sufficient dominance to make the other variant obsolete: Microsoft retains the repository name and a portion of the existing user base; AG2 retains the original founding team's continued attention and the strategic direction that motivated their departure. The resolution will require either Microsoft sunsetting its variant or AG2 establishing dominant share. Until then, the dual-lineage problem costs both variants share to LangGraph and CrewAI.

Why did OpenInterpreter lose share so sharply, and what does it indicate about the broader segment?

OpenInterpreter's technical positioning — a local-execution agent running a Python interpreter inside the agent loop — was attention-catching in the early agent moment but has been absorbed by the broader Claude Code, Cursor Composer, and Aider tooling. The project did not produce a commercial backing entity strong enough to sustain the engineering cadence required to compete with the better-funded alternatives, and founder Killian Lucas's attention moved toward a different product direction. The pattern is informative: founder-team driven open-source projects without commercial backing struggle to sustain the development cadence the segment now requires. The competition for developer attention has compressed substantially as enterprise buyers have moved toward production deployments, and the engineering investment required to remain a primary reference has risen accordingly.

How does the Block / Goose strategy differ from the LangChain Inc. or CrewAI Inc. strategies?

Block treats Goose as an open-source contribution that compounds Block's broader engineering platform investment, without depending on Goose as a primary revenue driver. LangChain Inc. and CrewAI Inc. depend on their respective open-source projects as the substrate that funnels enterprise customers into their commercial products (LangSmith, CrewAI Enterprise). The Block model is more stable in funding terms because the open-source project's investment is decoupled from its direct revenue contribution. The model is reproducible for any large engineering organisation with an internal platform investment that produces tooling worth releasing publicly. The structural lesson is that corporate-platform-team-led projects are more stable than founder-team or research-team-led projects because the funding model is less contingent on personal attention budget.

Why is the IDE-integration segment treated as distinct from the broader framework competition?

The IDE-integration segment — Continue, Cline, RooCode on the open-source side, Cursor and Windsurf on the closed-source side — addresses a substantively different problem than the orchestration frameworks (LangChain, LangGraph, CrewAI, AutoGen). The IDE-integration projects optimise for in-editor experience, autocomplete latency, codebase navigation, and integration with the developer's existing workflow. The orchestration frameworks optimise for production runtime, deployment patterns, observability, and multi-step workflow management. The two segments compete for different developer attention, address different problems, and have produced different competitive dynamics. Treating them as one segment obscures the structural reality that the broader framework projects have ceded the IDE-integration market to projects optimised for it.

What is the principal risk to LangChain Inc.'s present structural position?

LangChain Inc.'s commercial position depends on LangSmith remaining the category-defining agent observability product. Anthropic's deeper integration of agent capabilities and tooling into the Claude product surface — including its own observability and tracing capabilities — has the most potential to compress the category LangSmith operates in. If Anthropic's observability surface becomes sufficient for the production deployments that today require LangSmith, the funding model that underwrites LangGraph's open-source development weakens. The risk is not immediate; LangSmith's enterprise customer base is sticky and the product's feature breadth is meaningful. But the competitive trajectory is the principal risk to LangChain Inc.'s present position and warrants monitoring through the next twelve months.