Agentic Builders

FRAMEWORK RELEASES

LangGraph 1.2.0a6 tightens up v3 event streaming

via GitHub

If you've ever tried to debug a multi-agent graph by squinting at a wall of half-typed event payloads, this alpha is for you.

LangGraph dropped 1.2.0a6, an alpha that forwards kwargs cleanly through the stream_events(version='v3') protocol. It also bumps the notebook dependency from 7.5.2 to 7.5.6, which is the kind of housekeeping nobody notices until it breaks.

The v3 stream is the spine of how LangGraph surfaces what a workflow is actually doing in real time: every node entry, every tool call, every interim token. If you're piping those events into a tracer, a UI, or your own observability layer, the kwarg forwarding fix means fewer "where did my metadata go" tickets.

• Pre-release tagged on the langgraph repo, so pin explicitly if you pull it
• Improvement targets stream_events(version='v3'), not the older v1 / v2 streams
• Notebook bump is harmless but worth noting if you lock dev deps

Why it matters: Event streaming is where agentic frameworks actually earn their keep in production - if you can't see what the graph is doing, you can't ship it. Read more.

MCP IN THE WILD

An AI agent ran a full Criteo campaign for dentsu over MCP

MCP just graduated from "cool way to give Claude my filesystem" to "cool way to spend a media budget."

Criteo and dentsu say they ran the first fully orchestrated MCP ad campaign: an AI agent handled audience definition, launch, and optimization end-to-end through secure MCP API calls, with no human logging into a UI. What used to be a multi-day media planning workflow collapsed into a single conversational thread.

The target metric was completed views, so this isn't a synthetic demo. Real budget, real inventory, real performance reporting. That's a different bar than "agent successfully booked a flight on a sandbox."

• Agent talked to Criteo's stack purely over MCP, no browser automation hacks
• Replaces the traditional planner-driven workflow, not just augments it
• First public case of MCP carrying a revenue-generating business process

Why it matters: If you're building MCP servers, the design conversation just shifted from "what tools do we expose" to "what happens when an agent uses them with a credit card attached." Read more.

RUBRIC WATCH

A 32-point scoring system for "how autonomous is this agent, really"

Every team has a hand-rolled rubric for "is this thing actually agentic or is it a for-loop with vibes." Now there's a shared one.

The agentic.ai guide got a refresh on May 4 with a 32-point framework spanning eight dimensions: action capability, autonomy, planning, adaptation, state continuity, reliability, interoperability, and safety. Score 0 to 32, then map to one of six named levels from Reactive Tool up to Strategic Agent.

They've already run 272 tools through the framework, which gives you a reference set to argue with when your director asks where your shiny new orchestrator sits.

• Eight dimensions, four points each, deliberately boring math
• Six named levels so you stop reinventing tier names in every doc
• Explicit principle: higher autonomy is not the goal, fit-to-task is

Why it matters: Common vocabulary for agent capability beats yet another internal Notion page, especially when procurement starts asking what "Level 4" means. Read more.