ANW
Builder publicationAgent News Watch
Back to news
News

Roundups / Frameworks / Protocols

Weekly AI Agent Launch Roundup: ADK 2.0 Alpha, A2A 1.0, OpenAI Agents 0.13, and More

This week's AI agent launches for builders include Google ADK 2.0 alpha, A2A 1.0, OpenAI Agents 0.13, Mastra 1.14, CrewAI 1.11.1, and more.

Published

03/25/2026

Author

Agent News Watch

Lens

Launch context for teams deciding what changed and why it matters.

Composite of official release surfaces for Google ADK, A2A, OpenAI Agents, and CrewAI.
News brief

News coverage

Roundups / Frameworks / Protocols

Agent News Watch for teams building and operating AI agents.

This week's AI agent release pattern is clear: the market is moving away from vague agent positioning and toward more operationally explicit systems. The most important updates are not consumer demos. They are protocol stabilizations, graph runtimes, task APIs, MCP diagnostics, flow introspection, and session reliability fixes. For builders and operators, that is the signal worth tracking, because these are the changes that affect how real agent systems are wired, debugged, and governed.

For Agent News Watch readers, two guide questions sit underneath that release noise. One is which workflows are finally mature enough to deserve a real pilot, which is exactly the lens in AI Agent Use Cases. The other is whether orchestration and delegation updates actually justify specialist-role design, which is where Multi-Agent Architecture becomes the right follow-on read.

This week's biggest launch

Google's ADK 2.0 alpha is the clearest indicator of where framework vendors think production teams are headed. The release adds a graph-based workflow runtime and a Task API for structured agent-to-agent delegation, which means orchestration and multi-agent handoff are moving closer to first-class framework primitives instead of living in custom glue code. The important catch is that Google labels this release an early alpha with breaking changes to the agent API, event model, and session schema. In other words, it is a strong product signal, not an immediate migration order.

Launches by category

Frameworks and orchestration

Google ADK v2.0.0a1 (March 18)

Google used the first ADK 2.0 alpha to ship two substantive capabilities: a workflow runtime with routing, fan-out and fan-in, loops, retry, state management, dynamic nodes, human-in-the-loop, and nested workflows, plus a Task API for structured delegation between agents. For teams that have been bolting orchestration onto agent frameworks with custom graph layers, this is a meaningful move. Google is effectively saying that workflow control and agent-to-agent task handoff belong in the core framework surface. The alpha README reinforces that point by showing Workflow as a top-level primitive rather than an edge feature, which is why this release maps so cleanly into AI Agent Frameworks, AI Agent Architecture, and AI Agent Orchestration.

Builder takeaway: evaluate ADK 2.0 in a sandbox if you want graph execution and task delegation in one framework, but do not rush a production migration. Google explicitly warns that 1.x databases and sessions are incompatible with the alpha.

Mastra @mastra/core 1.14.0 (March 19)

Mastra's latest release is builder-relevant because it improves the operational layers that usually break first in production agents. The release adds AI Gateway tool support in the agentic loop, more reliable observational memory with getObservationsAsOf(), and per-server MCP diagnostics such as reconnectServer(serverName), listToolsetsWithErrors(), and getServerStderr(serverName). That combination matters because provider-executed tools, state replay, and MCP server failures are exactly the kinds of issues that turn a promising agent demo into a difficult on-call surface.

Builder takeaway: if you already run TypeScript agents with MCP or provider-executed tools, this is the kind of release that can reduce operator pain immediately, especially around debugging, replay, and runtime reliability.

LangGraph 1.1.3 (March 18)

LangGraph's latest release is smaller than ADK 2.0 alpha, but it still lands in the same strategic direction. The headline change is runtime execution info, alongside a checkpoint-postgres update. That is not glamorous, but it is exactly what graph-based agent teams need once flows run long enough to fail, pause, retry, or require review. Better execution data is what turns a graph runtime into something platform teams can support rather than just prototype.

Builder takeaway: if your team already treats graph orchestration as the control plane for agent runs, this release is more about observability and recoverability than new autonomy.

CrewAI 1.11.1 (March 23)

CrewAI's newest release adds flow_structure() serialization for Flow introspection and bundles a set of operator-facing fixes: preserved full LLM config across human-in-the-loop resume for non-OpenAI providers, a path traversal fix in FileWriterTool, and dependency updates addressing known security issues. The terse changelog undersells the practical impact. Flow introspection plus safer tool behavior is exactly the sort of maintenance work that matters when teams move from agent experiments to governed internal workflows. For teams turning those fixes into launch criteria, AI Agent Security and AI Agent Evaluation are the evergreen companions.

Builder takeaway: for teams running CrewAI in more controlled enterprise settings, this release is a useful combination of inspectability and risk reduction.

Protocols and interoperability

A2A v1.0.0 (March 12)

The A2A protocol hit 1.0.0 with a deeper set of changes than the version number alone suggests. The release includes breaking cleanup across push-notification config, request IDs, enums, and deprecated fields, plus new features such as a tasks/list method, multitenant gRPC support, SDK backward-compatibility support, and OAuth changes that remove implicit and password flows in favor of device code and PKCE. For builders who want agents built on different stacks to coordinate as agents instead of pretending every integration is just another tool call, that is a meaningful stabilization point.

Builder takeaway: if you are evaluating cross-agent interoperability, A2A just became much more concrete. The next step is not blind adoption. It is a compatibility test pass against the new shapes, auth flows, and task-management methods.

Agents and runtime tools

OpenAI Agents Python 0.13.0 (March 23)

OpenAI's latest Agents SDK release matters because it improves the boring parts that decide whether agent infrastructure is actually usable. The default Realtime websocket model now shifts to gpt-realtime-1.5, MCP servers gain list_resources(), list_resource_templates(), and read_resource(), and streamable HTTP MCP sessions now expose a resumable session_id. The release also fixes several runtime and session edge cases, including first-write races in SQLAlchemySession and tool-cleanup problems that can quietly destabilize long-running sessions.

Builder takeaway: this is most relevant for teams standardizing on MCP-backed agents or stateless worker patterns. Session resume support is especially useful if you are trying to scale agent workers without pinning every request to a single process.

OpenHands 1.5.0 (March 11)

OpenHands shipped a planning agent, a real-time task list tab, the ability to attach or change the Git repository for an existing conversation, and broader model support. For developer-agent teams, that mix matters because it makes the system easier to supervise. Planning visibility and explicit task status are not just UX improvements. They help engineers understand what the agent thinks it is doing before it changes code, and they make repo scoping less brittle in longer sessions.

Builder takeaway: if you are testing developer agents with human oversight, this release makes the system's work plan more inspectable instead of asking users to trust only the final output.

What changed technically this week

Three themes stand out across these releases.

First, orchestration is becoming more explicit. ADK adds a workflow runtime and structured task delegation. LangGraph adds runtime execution info. CrewAI adds flow introspection. The common thread is that framework authors know teams need inspectable run graphs and recoverable execution paths, not opaque prompt loops.

Second, interoperability is getting more operational. A2A 1.0.0 stabilizes more of the agent-to-agent contract, while OpenAI Agents and Mastra both improve MCP-facing behavior. That matters because agent systems fail at the seams: across runtimes, across remote resources, and across session boundaries.

Third, operator expectations are rising. Security fixes in CrewAI, better memory and diagnostics in Mastra, session-race fixes in OpenAI Agents, and visible planning in OpenHands all point to the same market reality: the next competitive edge is not just autonomy. It is operability.

What to watch next week

Watch for three follow-on signals. One is whether more frameworks start shipping first-class task delegation instead of leaving multi-agent handoff to custom code. Another is whether A2A 1.0.0 quickly triggers SDK and connector updates across agent platforms. The third is whether MCP tooling keeps shifting from it connects to it can be debugged, resumed, and governed under production conditions. For Agent News Watch readers, that is the real story line right now: the stack is getting less magical and more operable.

Keep reading

Use AI Agent Use Cases to decide which workflow now deserves a pilot, Multi-Agent Architecture when runtime and delegation updates point toward specialist roles, AI Agent Frameworks and AI Agent Orchestration to interpret the runtime and workflow-control launches, AI Agent Architecture for the system map, AI Agent Security for deployment controls, Model Context Protocol and Agent-to-Agent Protocol for interoperability choices, and AI Agent Evaluation for the production quality layer behind this week's release cycle.

For breakout coverage from this roundup, read the A2A v1.0.0 protocol brief and the Google ADK 2.0 alpha brief.

Sources

Google ADK v2.0.0a1 release

Google ADK 2.0 alpha README

A2A v1.0.0 release

A2A README

OpenAI Agents Python v0.13.0 release

Mastra @mastra/core 1.14.0 release

LangGraph 1.1.3 release

CrewAI 1.11.1 release

OpenHands 1.5.0 release

Turn this update into build work

Use these next reads to map the announcement into pilot fit, architecture, interoperability, or rollout controls.

Composite of IBM Think AI agent use case surfaces showing the guide title, introduction, and category list for support, research, operations, finance, and sales workflows.
Guide file

Guide coverage

Foundations / Implementation

Agent News Watch for teams building and operating AI agents.

Guide

Foundations / Implementation

Learn the best AI agent use cases for product, ops, engineering, and support teams, plus how to choose the right autonomy level, architecture, and rollout path.

Open guideRead more
LangChain supervisor diagram showing a coordinator agent routing work to specialist agents in a multi-agent workflow.
Guide file

Guide coverage

Architecture

Agent News Watch for teams building and operating AI agents.

Guide

Architecture

Learn when multi-agent architecture outperforms single-agent systems, which coordination patterns fit best, and how to manage context, reliability, security, and cost.

Open guideRead more
Composite of official framework surfaces from LangGraph, OpenAI Agents, Google ADK, and AutoGen.
Guide file

Guide coverage

Frameworks

Agent News Watch for teams building and operating AI agents.

Guide

Frameworks

Compare AI agent frameworks, understand when you need one, and learn how to choose the right stack for workflows, coding agents, and multi-agent systems.

Open guideRead more
Anthropic augmented LLM diagram showing model, retrieval, tools, and memory components.
Guide file

Guide coverage

Architecture

Agent News Watch for teams building and operating AI agents.

Guide

Architecture

Learn how AI agent architecture works across models, tools, memory, orchestration, guardrails, and multi-agent patterns with practical reference designs.

Open guideRead more
OWASP Top 10 for Large Language Model Applications project page with the GenAI security overview.
Guide file

Guide coverage

Security

Agent News Watch for teams building and operating AI agents.

Guide

Security

Learn how to secure AI agents against prompt injection, over-permissioned tools, unsafe memory, insecure handoffs, and risky outputs with practical controls.

Open guideRead more