Research Report · March 2026

State of AI Agents

A comprehensive analysis of 10 major agent frameworks, 4 interoperability protocols, memory architectures, and market dynamics — based on deep technical evaluation.

10
Frameworks Analyzed
$10.9B
Market Size 2026
97M
MCP Monthly Downloads
222K+
Combined GitHub Stars

The Agent Landscape Has Exploded

The AI agent ecosystem crossed a critical threshold in early 2026. Every major tech company now has a framework. Interoperability protocols (MCP, A2A) have moved from proposals to production. The question is no longer "should we build agents?" but "which stack?"

GitHub Stars by Framework
Protocol Support Matrix
Key Events — Q1 2026
January 2026
MCP + A2A join Linux Foundation AAIF
OpenAI, Anthropic, Google, Microsoft, AWS co-found the Agentic AI Foundation. Protocol governance becomes vendor-neutral.
February 2026
Microsoft Agent Framework RC
AutoGen + Semantic Kernel merge into unified Microsoft Agent Framework. GA targeted end of Q1.
March 2026
MCP V2 shipping, Karpathy autoresearch (53K stars)
MCP V2 with expanded capabilities. Karpathy's autonomous ML experimentation repo reaches 53K stars in 18 days. LangChain relaunches Open-SWE.
March 2026
NVIDIA Agent Toolkit at GTC
OpenShell announced with 17 enterprise partners. Enterprise agent deployments accelerate.
March 23, 2026
Dapr Agents v1.0 GA + Cisco DefenseClaw
CNCF-backed Dapr Agents hits 1.0 GA at KubeCon Europe (NVIDIA collab, ZEISS production). Cisco open-sources DefenseClaw — first agent security toolkit (MCP Scanner, A2A Scanner, AI BoM).
March 15, 2026
Deep Agents: 16.5K stars in 9 days
LangChain launches Deep Agents composable harness — fastest-growing new framework in history. AG2 ships ground-up Beta with streaming and AG-UI native support.
March 18, 2026
Google ADK 2.0 Alpha + Claude Opus 4.6
ADK 2.0 introduces graph-based workflows — biggest architectural change since launch. Claude Opus 4.6 launches with 1M token context. Microsoft Agent Framework GA imminent (merges AutoGen + Semantic Kernel). The agent infrastructure layer is consolidating.
April 2026
ICLR 2026 MemAgent Workshop
First dedicated academic workshop on agent memory systems (April 27). Signals academic recognition of memory as critical infrastructure.

10 Frameworks, Deep Technical Evaluation

Each framework was evaluated across architecture, memory, orchestration, protocol support, deployment options, and production readiness. Click column headers to sort. Click cards below for detailed analysis.

Framework Stars Architecture MCP A2A Languages License

The Protocol Stack Is Converging

Four protocols are forming the standard interoperability layer for AI agents. MCP leads adoption, A2A is maturing fast, and AG-UI / A2UI are emerging for frontend integration.

MCP
Model Context Protocol (Anthropic, 2024)
Maturity8 / 10
Monthly Downloads
97M+
Governance
Linux Foundation
SDK Languages
TS, Python
Next Release
V2, March 2026

The de facto standard for agent-to-tool communication. Every major framework now supports MCP natively or via adapters. Under Linux Foundation's AAIF with all major AI companies as co-founders.

A2A
Agent-to-Agent Protocol (Google, 2025)
Maturity6 / 10
Organizations
150+
Version
0.3
Transports
gRPC, HTTP, JSON-RPC
Blocker
No TS SDK

Horizontal agent-to-agent communication. Pre-1.0 but already deployed at Tyson Foods and Adobe. Huawei A2A-T (Agent-to-Agent for Telecom) launched at MWC 2026 — first sector-specific fork. Signed security cards and gRPC support in v0.3. Missing TypeScript SDK is a blocker for many teams.

AG-UI
Agent-User Interface Protocol
Maturity4 / 10
Status
Emerging
Key Adopter
CopilotKit, AG2
Supported By
LangGraph, ADK, AG2
Focus
Frontend streaming

Standardizes how agent backends stream state to frontend UIs. Now adopted by AG2 (v0.11.0, Feb 2026) in addition to LangGraph and ADK. Addresses a real gap — every team currently builds custom websocket/SSE bridges.

A2UI
Agent-to-UI (Google, 2025)
Maturity3 / 10
Status
v0.8 Spec
Creator
Google
Supported By
Google ADK
Focus
Generative UI

Google's spec for agents generating dynamic UI components. Standard component library integrated into ADK. Currently Google-only adoption — too early to evaluate ecosystem traction.

Protocol Support by Framework

Memory Is the Next Battleground

Agent memory determines whether agents can learn, adapt, and maintain context across sessions. Approaches range from simple checkpoint stores to sophisticated LLM-powered memory systems. The ICLR 2026 MemAgent Workshop (April 27) signals academic recognition of this critical gap.

Memory Capability Comparison

What We Learned

Eight deep dives, 4 protocol assessments, and 3 months of continuous monitoring distilled into the insights that matter most for teams building agent systems today.

Insight 01
Protocol convergence is real and accelerating
MCP + A2A under Linux Foundation governance means the protocol stack is no longer a bet — it's becoming infrastructure. Teams that don't adopt MCP by mid-2026 will be excluded from the emerging tool ecosystem. 97M monthly downloads speaks louder than any specification.
Insight 02
No framework wins on all axes — pick your tradeoff
LangGraph leads production readiness. Google ADK leads protocol completeness (all 4 protocols). CrewAI leads approachability. OpenAI Agents SDK leads simplicity. AutoGen leads community size. The "best" framework depends entirely on your constraints.
Insight 03
Memory systems are immature across the board
No framework has production-grade long-term memory. CrewAI's unified memory is the most sophisticated but unproven at scale. Every framework either lacks cross-session persistence, semantic search, or memory pruning. This is where the most research investment is needed (confirmed by ICLR 2026 MemAgent Workshop).
Insight 04
Agents are great at bugs, bad at features
FeatureBench (March 2026) shows agents solve 74% of bug tasks but only 11% of feature tasks. SWE-Skills-Bench found only 7/49 agent skills actually help in practice. The gap between agent hype and capability is widest on creative, multi-step feature work.
Insight 05
The organizational orchestration layer is wide open
Every framework focuses on task-level orchestration (run this workflow). Nobody addresses organizational orchestration (persistent agents with identity, career-spanning memory, human-org integration). This is a genuine whitespace. The closest approach is file-based memory validated by ICLR 2026's "Codified Context" research.
Insight 06
Autonomous loops are the new paradigm
Karpathy's autoresearch (53K stars in 18 days) proved that autonomous experiment loops work at scale — 700 experiments, 20 improvements, found bugs in hand-tuned code. Shopify independently validated: 53% faster rendering, 61% fewer memory allocations. The "Loopy Era" is here.
Insight 07
Enterprise fragmentation creates lock-in risk
Microsoft is merging 3 frameworks (AutoGen + Semantic Kernel + Agent Framework). Google's ADK is best on GCP. Every vendor's "open source" framework works best with their cloud. True vendor-neutral production deployment remains difficult.
Insight 08
Security is an afterthought everywhere
The Corba vulnerability showed 79-100% of AutoGen agents can be blocked in 1.6-1.9 turns. No framework has built-in governance, audit trails, or execution-aware safety. As agents gain more autonomy, this gap becomes critical. Agent security is 2016-era web security.

Which Framework Should You Use?

Answer 5 quick questions about your use case, team, and priorities — we'll recommend the best frameworks for your specific situation based on our deep evaluation of all 10 frameworks.


Head-to-Head Framework Comparison

Select 2–4 frameworks to compare them side by side across production readiness, simplicity, protocol support, community, memory, and enterprise capabilities.

Select 2–4 frameworks to compare
Pick frameworks above to see them compared
Radar chart + feature-by-feature breakdown

How This Report Was Produced

This report was produced by an AI research team using a multi-agent orchestration system. Each framework analysis involved source code review, documentation study, community sentiment analysis, and cross-referencing with academic publications.

01
Source Identification
GitHub repositories, official documentation, release notes, changelogs, blog posts, and academic papers for each framework. Community signals from Hacker News, Reddit, and Discord.
02
Deep Technical Evaluation
Each framework evaluated across 6 dimensions: architecture, memory, orchestration, protocol support, deployment, and community. Verified claims against actual source code and issues.
03
Cross-Reference & Synthesis
Findings cross-referenced across frameworks to identify patterns, gaps, and genuine differentiators. Protocol maturity assessed via SDK downloads, enterprise adoption, and specification completeness.
04
Insight Extraction
Strategic insights derived from patterns across all evaluations. Validated against market data (Gartner, MarketsAndMarkets), academic research (ICLR 2026, IMWUT), and enterprise case studies.

About the research team. This report was produced by RyanHub's AI Research Lead as part of an ongoing landscape monitoring program. All data points are sourced from public information as of March 24, 2026. Framework evaluations represent a snapshot in time — the agent ecosystem moves fast.

Limitations. Benchmarks are self-reported by framework authors unless otherwise noted. GitHub stars are an imperfect proxy for adoption. Enterprise deployment numbers are typically not public. We do not evaluate cost (API pricing varies too much by use case). Memory evaluations are based on documented capabilities, not load testing.

Updates. This report is updated on a weekly cadence. For corrections or additions, open an issue on GitHub.