The short version: In one quarter, the agent-framework field went from moving targets to stable APIs. Microsoft Agent Framework hit 1.0 on April 3, Pydantic AI hit V2 on June 23, LlamaIndex Workflows and CrewAI both shipped maturity milestones, and the Model Context Protocol's 2026-07-28 release candidate rips out sessions to make tool servers stateless. If you were waiting for the ground to stop shifting before committing to a stack, this was the quarter. Here is each release in two lines.
The pattern this quarter is not new features. It is stability promises. Frameworks are finally betting their reputations on APIs you can build a company on.
Microsoft Agent Framework 1.0 (April 3, 2026)#
What shipped: The GA release that merges Semantic Kernel and AutoGen into one open-source SDK for .NET and Python, with native MCP (tool discovery) and A2A (agent-to-agent) support, six model providers, and declarative YAML agent definitions.
What it means for a founder: If your stack is .NET, you finally have a first-party agent framework with a long-term-support commitment instead of two competing research projects. The A2A support matters if you plan to let your agents talk to agents you do not own. See our three-thresholds comparison before you commit.
Anthropic Claude Agent SDK#
What shipped: The renamed (formerly Claude Code) SDK now runs subagents in the background by default, each with its own isolated context window, and lets a lead agent plan work and then fan out to many parallel subagents in a single session.
What it means for a founder: Isolated-context subagents are the cheapest way to keep a long agent run from drowning in its own history. If your product does research, migrations, or anything that decomposes into parallel chunks, this is the pattern that keeps context windows from becoming your cost ceiling.
CrewAI 1.14 (June 2026)#
What shipped: Pluggable default backends for memory, knowledge, RAG, and flow, a Chat API for conversational flows, native Snowflake Cortex support, and richer LLM event data (real finish_reason, sampling params, response.id).
What it means for a founder: Pluggable backends mean you are no longer stuck with CrewAI's built-in memory store when you outgrow it. You can point memory and RAG at your own infrastructure without leaving the framework. Details in our pluggable-memory deep-dive.
Pydantic AI V2 (June 23, 2026)#
What shipped: A harness-first redesign built on one primitive: the capability. Tools, instructions, lifecycle hooks, and model settings all collapse into a single composable unit. Core stays small; the first-party Harness (memory, guardrails, code mode, tool search) ships the batteries.
What it means for a founder: One primitive means less framework to learn and less glue code to maintain. If you value type-safety and a small stable core you can reason about, V2 is the cleanest mental model shipped this quarter. Read why capabilities change the shape of your agent code.
LlamaIndex Workflows 1.0 (June 2026)#
What shipped: The first stable, standalone release of Workflows, an event-driven agent framework for Python and TypeScript, with typed workflow state on both runtimes and resource injection (database clients and more) for Python.
What it means for a founder: Workflows split into its own package and repo, so you can adopt the orchestration layer without pulling in all of LlamaIndex. Typed state across Python and TypeScript is the practical win if your backend and frontend both touch the agent loop.
LangGraph 1.2 line (Q2 2026)#
What shipped: Per-node timeouts (timeout= on add_node, plus TimeoutPolicy for wall-clock and idle limits, raising NodeTimeoutError and handing off to your retry policy) and type-safe streaming v2, a unified StreamPart output with typed chunks. Async nodes only.
What it means for a founder: Per-node timeouts are the unglamorous feature that keeps one slow tool call from hanging a whole run in production. If you already run LangGraph, this is a reliability upgrade you want before your next launch.
MCP 2026-07-28 release candidate (stateless core)#
What shipped: The largest revision of the protocol since launch. It removes protocol-level sessions (no more initialize handshake or Mcp-Session-Id), moving protocol version and client info into _meta on every request, and adding response caching (ttlMs, cacheScope) so any server instance can answer any call. It also deprecates sampling, roots, and logging, which remain functional for at least twelve months.
What it means for a founder: This is the big one. Stateless means your MCP tool servers scale horizontally behind an ordinary load balancer, no sticky sessions, no shared session store. That is a real infrastructure cost saving. But read the deprecations before you build: if you rely on server-initiated sampling, you have a migration ahead. Our full breakdown is in MCP goes stateless.
What to actually do this week#
If you are starting fresh, pick the framework that matches your language and support needs; they are all at stable-API milestones now, so the switching cost of a wrong guess is lower than it was in Q1. If you already ship on MCP, audit your servers for sampling, roots, and logging before the clock on those twelve months starts. And if you build inside an editor, the VS Code July agent update is worth a look. The ground stopped moving this quarter. Build on it.
