Chapter 1. The 2026 Landscape: Frameworks and Orchestrators

By mid-2026 the agent framework market had gone through a hard consolidation: every major lab is down to exactly one "blessed" framework, the losing projects have been officially shut down or moved to maintenance mode, and the agent interoperability protocols (MCP, A2A) have passed under Linux Foundation governance. Yet the ecosystem's maturity is deceptive: two of the three most-downloaded SDKs still have no version 1.0, and analysts simultaneously predict both explosive growth and the cancellation of more than 40% of agentic projects. Choosing a framework in 2026 is first and foremost a bet on the vendor's strategy, not on the current feature set.

The Great Consolidation of H1 2026

The first half of 2026 cemented the "one lab, one framework" principle. Microsoft merged Semantic Kernel and AutoGen into a single Microsoft Agent Framework: public preview on October 1, 2025, Release Candidate on February 19, 2026, and 1.0 GA on April 3, 2026 — the date confirmed by Microsoft's official announcement. Both predecessors have been moved to maintenance (an important fact-checking caveat: the words "maintenance mode" never appear in the GA post itself — that status is recorded on Microsoft Learn and in third-party reviews). The microsoft/autogen repository, 59.7K stars and all, is effectively frozen; the surviving community fork AG2 exists, but at niche scale — 555K downloads a month against CrewAI's 11.4M (a breakdown of the fork).

OpenAI walked the same path: Swarm was long ago demoted to a teaching example, and on June 3–4, 2026, deprecation notices went out for the visual Agent Builder and the Evals platform, with a shutdown date of November 30, 2026 (the notice in the OpenAI community). A nuance the retellings tend to lose: the migration path points not only to the code-first Agents SDK but also to the no-code ChatGPT Workspace Agents, and ChatKit is not deprecated — so this is only a partial retreat from low-code, not a full reversal.

graph LR
    SK[Semantic Kernel] --> MAF[Microsoft Agent Framework 1.0]
    AG[AutoGen] --> MAF
    AG --> AG2[AG2 fork, niche]
    SW[OpenAI Swarm] --> SDK[OpenAI Agents SDK]
    AB[Agent Builder] -- shutdown 2026-11-30 --> SDK
    AB -- partially --> WS[ChatGPT Workspace Agents]
    CCSDK[Claude Code SDK] --> CAS[Claude Agent SDK]

Who Weighs What: Live Numbers as of July 12, 2026

The R01-frameworks dossier relies on direct queries to the PyPI stats API, the npm API, and the GitHub API as of July 12, 2026 (not on blog-post retellings), and every number below has passed independent re-verification.

LangGraph is the download king and the enterprise default. Its 1.0 GA on October 22, 2025 came with a "no breaking changes until 2.0" promise and a roster of production users: Uber, LinkedIn, Klarna, JPMorgan, BlackRock, Cisco (the LangChain announcement). Anthropic's Claude Agent SDK is the fastest-growing: 22M downloads a month on version 0.2.x with a mere 7.6K stars — growth driven by real usage, not GitHub hype. In May 2026, at "Code with Claude," Anthropic showed dynamic workflows (a lead agent hands tasks to parallel subagents on a shared file system) and "Outcomes" rubric grading; the official post claims only a success-rate lift "of up to 10 points on the hardest tasks" (Anthropic) — the +8.4%/+10.1% figures for Word/PowerPoint floating around the reviews exist in exactly one secondary retelling and must not be cited as fact.

Two caveats to the table are mandatory. First, download counts tally CI pipelines and transitive dependencies just as readily as humans: LangGraph estimates vary by up to 2x between sources, and PydanticAI's numbers are inflated by transitive installs of pydantic-ai-slim. Second, a "GA" label guarantees no stability: the No. 2 and No. 3 entries in the ranking live on 0.x, and OpenAI shut down its own eight-month-old Agent Builder — only LangGraph has a semver promise.

Protocols and the Low-Code Flank

In 2026 the interoperability layer became neutral territory. MCP (introduced in November 2024) was handed over by Anthropic on December 9, 2025 to the Agentic AI Foundation — a foundation under the Linux Foundation umbrella, co-founded by Anthropic, Block, and OpenAI — with 10,000+ active public MCP servers and 97M+ monthly SDK downloads (the Anthropic announcement). Per a Stacklok survey, 41% of software organizations already run MCP in limited or broad production. A2A was transferred by Google to the Linux Foundation in June 2025; gRPC and signed Agent Cards arrived in v0.3, by 2026 A2A had reached v1.0.0, the ecosystem spans over 150 organizations, and native support ships in ADK, LangGraph, CrewAI, LlamaIndex, and Agent Framework (Google Cloud).

The low-code flank is dominated by n8n: on May 12, 2026, a strategic investment from SAP doubled its valuation to $5.2B, the platform is being embedded into SAP Joule Studio, and it claims 1,400+ enterprise customers and 1.7M monthly active builders (PR Newswire). n8n's 196K stars are roughly 3.3–3.5x the starriest code-first frameworks. The TypeScript niche belongs to Mastra: 1.0 in January 2026, a $22M Series A from Spark Capital in April (the Mastra announcement), and growth from 300K to 1.11M weekly downloads in six months.

The Layer Below the Frameworks: Harness Orchestrators and Agent Catalogs

Between the SDKs and day-to-day practice a separate layer has grown up — coding-agent orchestrators and harness libraries, nearly invisible in PyPI statistics. The map of the layer is the catalog-index awesome-agent-orchestrators (created 2026-01-20; 1,040 stars as of July 18, 2026), but it must be used as a discovery source, not a registry of facts: the entries are authors' self-descriptions submitted via PR (the maintainer lists his own amux), and the tool count is volatile on a scale of days — one count yields 135 entries, a repeat gives ~155, and the category breakdown does not reproduce. Only a dated snapshot is usable.

The poles of the layer are telling. The maximalist: ruflo (ruvnet/ruflo, ex-claude-flow — "Claude Flow is now Ruflo"; 64,913 stars, MIT, release v3.32.4 of 2026-07-17) — a "Queen-led hive-mind" with Raft/Byzantine/Gossip consensus, ~210 MCP tools, 100+ agents; README claims like "89% routing accuracy" and "8.1M+ ecosystem downloads" have been audited by no one — a showcase of the niche's ambitions, not a verified benchmark. The pragmatist: OpenWiki (langchain-ai, MIT) — a CLI agent that documents code and a personal wiki; created 2026-06-22, it collected 12,181 stars and 839 forks in ~4 weeks (v0.2.0 of 2026-07-16) — organic growth of rare speed in the niche of agent-maintained documentation. Alongside them sits the agent-loop library loopy (Forward-Future / Matthew Berman, MIT; 2,747 stars, commit of 2026-07-07): the living catalog has grown to 85 loops in 5 categories — engineering, evaluation, operations, content, design (loop-library). And the meta-fact of the era: the awesome-harness-engineering index (ai-boost, CC0 1.0; 3,117 stars, commit of 2026-07-17) is partly maintained by the agent itself — of the 10 most recent commits, ~4 are authored by "Claude Code." A list about harness engineering that is maintained by a harness is the exact self-portrait of 2026. Every number in this section is a GitHub API snapshot as of July 18, 2026: in this layer they drift daily.

The Market: Peak Hype and the Trough at Once

The R12-market dossier paints a picture most honestly described as a managed contradiction. In June 2025 Gartner predicted the cancellation of more than 40% of agentic projects by the end of 2027 and named the phenomenon "agent washing": of the thousands of vendors claiming agentic products, roughly 130 have real capabilities. Two months later — same analyst, same Gartner — the forecast: 40% of enterprise applications will feature task-specific agents by the end of 2026 (versus <5% in 2025), and $450B+ in agentic revenue by 2035. The two theses are compatible: mass shallow embedding of agents by vendors, plus high mortality among ambitious autonomous projects. The demand side is more sober: per Gartner's 2026 CIO survey, only 17% of organizations have actually deployed agents (against 60%+ stated intent over two years); per McKinsey, 62% are at least experimenting, but no more than ~10% have scaled agents within any single function. The famous "95% of pilots with no P&L effect" from MIT NANDA is methodologically weak (a small, self-selected sample) — use it as a direction, not a precise figure.

The money, meanwhile, flows to where agents already work: Cursor has grown to a ~$4B annualized run rate and a $60B deal with SpaceX (TechCrunch), and Cognition raised ~$1B at a ~$26B post-money valuation (about $25B pre-money) on a $492M run rate (TechCrunch). A correction applies here too: a run rate is an extrapolation, not audited revenue, and the margins of coding agents are contested because of inference costs.

What Download Numbers Don't Show

Three things a practitioner must know before taking the table above to leadership. First: framework fatigue is real — a 2026 field guide estimates that about half of engineering teams write agents in plain Python with no framework at all, and a popular community position holds that ~90% of "agentic" projects would be better served by simple prompt chains (SocialCrawl). Second: the base of "known production users" is thin — Klarna appears both as LangGraph's showcase case (85M users, an 80% cut in time to resolution) and in OpenAI AgentKit materials; the same logos circulate across every vendor. Third: vendor metrics are unaudited — CrewAI's claim of "60%+ of the Fortune 500" sits next to $18M in total funding and revenue estimates of ~$3M; the typical pattern seen in practice is a prototype on CrewAI, then a migration to LangGraph for production state management.

What to Apply Tomorrow

1. For a new production project in Python, take LangGraph by default: it is the only framework with an explicit "no breaking changes until 2.0" semver promise and the longest roster of confirmed enterprise users; on a TypeScript stack, take Mastra.
2. Before adopting any framework at all, check whether a prompt chain is enough: about half of teams get by with plain Python, and for linear workflows that is cheaper to maintain than any orchestrator.
3. Build in MCP as your tool layer from day one — the protocol is neutral (Linux Foundation), backed by every lab, and already in production at 41% of software organizations; it is your main insurance against vendor lock-in.
4. Pin versions for everything younger than 1.0 (OpenAI Agents SDK 0.18.x, Claude Agent SDK 0.2.x): the second and third most popular SDKs offer no compatibility guarantees, and the Agent Builder shutdown precedent shows that even flagship products can get deprecated.
5. Filter vendor numbers when defending a budget: separate live metrics (PyPI/npm downloads — with a ±2x correction for CI), audited revenue, and marketing claims like "60% of the Fortune 500"; the ready answer to Gartner's 40%-cancellation forecast is that the same Gartner expects 40% of applications to feature agents by the end of 2026.