Agent Frameworks

Executive Summary

The AI agent framework market expanded sharply in 2025-2026. The independent leaders — LangChain/LangGraph, CrewAI, LlamaIndex, Mastra, Pydantic AI, Agno — now compete directly with first-party frameworks from the foundation labs and hyperscalers: the OpenAI Agents SDK, Google ADK, and Microsoft Agent Framework. This report covers eleven frameworks across Python, TypeScript, and .NET.

Key Findings:

• LangChain still dominates — 100M+ monthly downloads, 139k+ GitHub stars, and a $1.25B valuation after its October 2025 Series B^[1][2]
• The labs moved in — OpenAI (27k+ stars), Google (20k+), and Microsoft (11k+) all ship open-source first-party frameworks^[3][4][5]
• AutoGen is superseded — Microsoft Agent Framework hit 1.0 GA in April 2026; AutoGen is community-maintained with no new features^[6]
• TypeScript surged — Vercel AI SDK downloads nearly tripled to 57M+/month, and Mastra hit v1.0 with a $22M Series A^[7][8]
• Specialization persists — LlamaIndex for document AI, CrewAI for multi-agent business automation, Pydantic AI for type-safe durable execution, Agno for self-hosted runtimes

Strategic Planning Assumptions:

• By 2027, 80% of production agent deployments will require integrated observability platforms
• By 2028, first-party lab frameworks will power 40% of new agent projects, pressuring independents toward neutrality and multi-cloud as differentiators
• By 2028, TypeScript agent frameworks will capture 30% market share (up from ~15% today)

---

Market Definition

Agent frameworks are software libraries that provide abstractions for building AI agents — autonomous systems that use LLMs to reason, plan, and execute multi-step tasks.

Inclusion Criteria:

• Provides agent orchestration primitives (not just LLM wrappers)
• Open source core (MIT, Apache, or equivalent)
• Active development with production users
• Multi-model support (not locked to single provider)

Exclusion Criteria:

• Pure LLM API wrappers without orchestration
• Proprietary-only frameworks
• Abandoned or pre-alpha projects
• Single-provider solutions — notably the Claude Agent SDK (strong traction, but Claude-only)

Adjacent, not included: Smolagents (Hugging Face's deliberately minimal code-first agents), Letta (a stateful agent server, not a library), and LiveKit Agents (realtime voice — a separate category).

---

Comparison Matrix

---

Product Profiles

Agno

Highest-starred independent framework with a self-hosted production runtime^[9]

• 40k+ GitHub stars (June 2026) — the most-starred framework in this comparison
• AgentOS runtime — FastAPI control plane with RBAC, scheduling, and chat surfaces, run in your own VPC
• Model-agnostic — 30+ providers; rebranded from Phidata in January 2025
• Lightly funded — $5.4M disclosed; commercial layer is young

---

AutoGen

The multi-agent pioneer, now superseded by Microsoft Agent Framework^[6]

• Maintenance mode — community-managed, no new features; last feature release September 2025
• 58.9k stars remain, but Microsoft's own guidance points new projects to its successor
• Migration path — official guide maps AutoGen patterns onto Microsoft Agent Framework

---

AWS Strands Agents

Amazon's model-driven SDK — quiet stars, enormous downloads

• Model-driven loop — the LLM plans its own steps; graph/swarm patterns for multi-agent
• Production inside Amazon — powers Amazon Q Developer, AWS Glue, VPC Reachability Analyzer
• 16.7M monthly PyPI downloads against only 6.1K stars — distribution via AWS, not community
• Repositioning — core repo renamed to harness-sdk under Bedrock AgentCore's harness preview

---

Cargo AI

Auditable AI tools as compiled Rust binaries — the smallest project in this comparison^[10]

• JSON-first definitions — Declare inputs, schema, and actions once; compile to a native executable for macOS/Linux/Windows
• Auditability as the thesis — Definitions are diffable and reviewable; the binary is the artifact
• Model-flexible — OpenAI (including Codex auth), Ollama, and direct API key paths
• Very early — 4 stars, solo developer, bus factor of one; the cargo-ai.org registry is live but unproven

---

CrewAI

Production multi-agent automation — 63% of Fortune 500^[11]

• 63% Fortune 500 adoption, 450M+ workflows monthly, 53k+ stars (June 2026)
• Crews + Flows — autonomous agent teams plus deterministic workflows
• Pricing simplified — free tier (with tracing) + custom Enterprise; the public Professional tier is gone
• Still Python-only — no first-party TypeScript SDK shipped

---

Google ADK

Google's code-first framework with five language SDKs and a GCP deploy path^[4]

• Hybrid orchestration — deterministic workflow agents plus LLM-routed hierarchical delegation
• 20k+ stars (Python), with TS, Go, Java, and Kotlin SDKs
• Enterprise-grade path — Vertex AI Agent Engine for managed sessions, memory, and observability
• A2A native — Google's agent-to-agent protocol built in

---

LangChain/LangGraph

#1 downloaded framework — now a $1.25B company with 1.0 stability^[1][2]

• 100M+ monthly downloads, 139k+ stars, 6k+ LangSmith customers
• 1.0 GA (October 2025) for both LangChain and LangGraph — API stability after years of churn
• Platform expansion — LangSmith Engine, Sandboxes, Fleet, and an LLM Gateway shipped at Interrupt 2026
• $125M Series B at $1.25B valuation (IVP, October 2025)

---

LlamaIndex

Document AI leader — 1B+ documents processed, Workflows 1.0 stable^[12]

• 1B+ documents processed; 300k+ LlamaParse users; ~50k stars
• Workflows 1.0 stable in Python and TypeScript; LlamaAgents in open preview
• LiteParse — new open-source local parser softens the LlamaCloud dependency
• Pricing increased — Starter $50/mo, Pro $500/mo (from $29/$299)

---

Mastra

TypeScript-native framework — v1.0 stable with a $22M Series A^[8]

• v1.0 stable (January 2026) with API stability guarantees; ~25k stars
• $35M total raised — $22M Series A led by Spark Capital (April 2026)
• Hosted platform — Studio, Server, and Memory Gateway all GA
• Observational Memory — 94.87% on LongMemEval remains the highest published full-benchmark score

---

Microsoft Agent Framework

The unified successor to AutoGen and Semantic Kernel — 1.0 GA April 2026^[5][13]

• .NET + Python with API parity; graph-based workflows with checkpointing
• Five GA orchestration patterns — sequential, concurrent, handoff, group chat, Magentic-One
• OpenTelemetry-native distributed tracing; pluggable memory (Mem0, Redis, Neo4j)
• 11k+ stars and weekly releases since GA

---

OpenAI Agents SDK

OpenAI's lightweight primitives — the most-adopted lab framework^[3]

• 27k+ stars (Python) plus a JS/TS SDK; ~weekly releases
• Three primitives — Agents, Handoffs, Guardrails — plus Sessions and built-in tracing
• Multi-model via LiteLLM (100+ providers), but OpenAI-first by design; tracing defaults to OpenAI's dashboard
• Still v0.x despite GA status since March 2025

---

Pydantic AI

Type-safe agents with durable execution, from the Pydantic team^[14]

• Validated structured outputs across any provider; pydantic-graph control flow
• Durable execution — first-class Temporal, DBOS, Prefect, and Restate support
• Logfire observability — OpenTelemetry-based, from the same company
• 17.7k stars; v1 stable with a no-breaking-changes policy, v2 in beta; $17.2M raised (Sequoia-led Series A)

---

Vercel AI SDK

The TypeScript standard — downloads nearly tripled to 57M+/month^[7]

• 57M+ monthly downloads (June 2026, up from 20M at SDK 6 launch); 24.8k stars
• SDK 6 GA — ToolLoopAgent, tool approval, full MCP support; SDK 7 in canary
• AI Gateway — 100+ models at list price, no token markup
• Workflow DevKit — durable agents and secure sandboxed code execution

---

Architecture/Pattern Analysis

Orchestration Approaches

Durability and Memory

---

Gap Analysis

Bottom line: All eleven support MCP and human-in-the-loop. The differentiators are language stack, durable execution depth, and whether observability requires the vendor's cloud.

---

Strategic Recommendations

By Use Case

By Buyer Profile

Enterprise with complex documents: → LlamaIndex + LlamaCloud for parsing; LangChain for broader agent needs

Fortune 500 automating business processes: → CrewAI with AMP — proven at scale with enterprise support

TypeScript-first team: → Vercel AI SDK for the standard toolkit; Mastra for batteries-included framework + memory

Cloud-standardized shops: → Match your platform: OpenAI Agents SDK, Google ADK, or Microsoft Agent Framework. All are open source, but each pulls toward its vendor's observability and deployment stack

Startup needing fastest time-to-production: → LangChain + LangSmith — largest community and examples; Pydantic AI if type safety and durability matter more than breadth

---

Market Outlook

Near-Term (2026-2027)

• First-party frameworks (OpenAI, Google, Microsoft) take share among single-cloud shops; independents respond by deepening neutrality and durability features
• LangChain consolidates its platform play (Engine, Sandboxes, Gateway) post-Series B
• CrewAI's next funding round (none since 2024) is a watch item

Medium-Term (2027-2028)

• Durable execution becomes table stakes — Temporal-style guarantees expected in every major framework
• Observability consolidates around OpenTelemetry; vendor-locked tracing becomes a liability
• TypeScript frameworks capture significant share as agent UIs and full-stack apps converge

Long-Term (2028+)

• Agent frameworks become infrastructure layer (like databases)
• Open standards (MCP, A2A) enable framework interoperability — already shipping in ADK, MAF, and Pydantic AI
• Consolidation hits the middle: leaders and first-party frameworks survive; undifferentiated mid-tier projects fade

---

Bottom Line

Current Leaders:

• LangChain/LangGraph for ecosystem breadth, observability, and now stability (1.0)
• CrewAI for production multi-agent business automation
• LlamaIndex for document understanding and RAG

Rising Forces:

• The lab frameworks — OpenAI Agents SDK, Google ADK, and Microsoft Agent Framework make "use your cloud vendor's framework" a defensible default
• Pydantic AI and Vercel AI SDK — type safety and durable execution are the new differentiation axes
• Agno — the star-count leader betting on self-hosted runtimes

The market expanded rather than consolidated in 2026. Choose based on primary use case, language stack, cloud alignment, and whether observability must stay in your infrastructure. AutoGen is the one framework on this list that should not be chosen for new projects — Microsoft says so itself.

---

Research by Ry Walker Research • methodology