Each layer of the agent toolchain emits a different kind of signal right now. Reading those signals carefully is worth more attention than most people give them.

Start where things are quietest. Playwright has 84,000 GitHub stars, roughly 37 million weekly npm downloads, and a 45% adoption rate among surveyed QA teams. Selenium, the incumbent for over a decade, sits at 22%. Microsoft is actively extending Playwright to serve agentic workflows via a new CLI designed specifically for AI coding agents. When the replacement tool has doubled the adoption of the thing it replaced and the platform owner is investing in its next use case, the category looks settled enough to build on. That's the simplest consolidation signal to read.

Settled categories create pressure on the layers above them, though. Stagehand v3, shipped in February 2026, is a complete architectural rewrite of an AI-native browser automation tool that was previously built on top of Playwright. The new version drops the Playwright dependency entirely, talking directly to Chrome DevTools Protocol instead. The reasoning is specific: Playwright's auto-waiting and actionability checks, designed for testing, added overhead when the real job was streaming raw accessibility trees and DOM snapshots to a model. So the team went around the dominant layer rather than continuing to build on top of it.

A different kind of signal. Call it architectural divergence. Browser automation for testing and browser automation for agents may not share a foundation for long.

Protocol work looks different still. WebMCP landed as a W3C Draft Community Group Report in February 2026, currently behind a flag in Chrome Canary. Google, Microsoft, Mozilla, and Apple are all at the table. The concept: websites expose structured tools that agents can call natively in the browser, through both a Declarative API for standard HTML forms and an Imperative API for dynamic JavaScript interactions. The page itself becomes the tool provider. Four browser vendors participating before there's a shipping implementation looks like institutional construction. Committee participation and shipped standards are separated by years of negotiation, so this is the earliest and most uncertain kind of consolidation signal. But it tells you the layer is being deliberately assembled by committee, with coordination preceding competition.

The orchestration layer looks nothing like any of these. Every major lab now ships its own opinionated SDK into a space independent frameworks already occupy. In March 2026 alone, five companies shipped agent frameworks. LangChain still holds 126,000 GitHub stars and LangSmith has processed over 15 billion traces. But OpenAI, Google, and Anthropic are all betting that tight model integration outweighs ecosystem breadth. The opening phase of a land grab, with no clear winner in sight.

Layer	Example	Signal Type
Browser automation	Playwright	Numerical dominance
AI-native abstraction	Stagehand v3	Architectural divergence
Protocol standardization	WebMCP	Institutional construction
Orchestration / SDKs	Lab SDKs vs. LangGraph, CrewAI	Active contest

Four layers, four signal types. The bottom of the stack appears stable enough to commit to. The middle is splitting along a paradigm boundary that hasn't fully resolved. The top is where the ground is still shifting. Which kind of movement you're looking at shapes the decision in front of you.