The Agent Framework Race Is Now About Protocols, Not Demos

The protocol layer

Google has launched A2UI 1.0 as a stable, framework-agnostic open standard for generative UI, according to CopilotKit, which has built its Generative UI framework on the protocol. A2UI 1.0 supports bidirectional function calls between the UI and the agent, lets a single screen combine multiple design systems and component libraries, and allows a full UI to be created in a single message. The protocol is data-only and does not execute code, which CopilotKit says prevents agents from inventing unauthorized UI elements, and its catalogs include composition rules defining allowed parents and children for components. A2UI 1.0 natively travels over the AG-UI protocol, and supports official renderers for React, Angular, Lit, Flutter, SwiftUI, and Jetpack Compose. CopilotKit says its Generative UI includes a catalog of 70 ready-made components. m:…8299

The protocol matters because it answers a question every agent builder hits: who decides what the user sees. With A2UI 1.0, validation results include severity levels of error, warning, or info, and functions with side effects require user activation. Components can point to their own specific catalogs, and multiple catalogs can sit on a single surface. The design pushes the agent's freedom into a bounded space: the model can compose, but only from approved parts, and only under rules the catalog enforces. m:…8299

Selling the engineer, not the tool

Mastra is fighting the same race from the opposite side. The company says it has launched Mastra Forward, an AI engineering-as-a-service offering that provides a named forward-deployed engineer who works in the client's codebase, backed by the rest of Mastra's core engineering team. The service includes Mastra Factory, installed in the client's repository to run a swarm of agents, plus an enterprise license for hosted, self-hosted, or bring-your-own-cloud deployment, a Trace Intelligence pool with evals run as a service, and a quarterly architecture review. Mastra says code produced by its engineers belongs to the client, and that improvements to the framework, Factory, Harness, or Knowledge are integrated into the core. The engagement is annual, invoiced quarterly or prepaid upfront, the company says, and Mastra claims the service costs less than one senior AI hire. m:…0229

Mastra's pitch rests on a production claim: the company says it uses its software factory to produce over 750 pull requests per week. The argument is that hiring is the bottleneck, not the model, and that the fastest way to ship agents is to rent the team that builds the framework. Whether that holds depends on something the announcement does not show: how much of that pull-request volume is client work, and how much is the factory's own output. m:…0229

Everyone builds now

n8n argues the bottleneck is organizational, not technical. The company says many organizations currently rely on a small, specialized team to build automation on behalf of the rest of the company, and that centralized automation teams often become bottlenecks as a company scales. AI assistants like n8n and MCP have significantly lowered the technical barrier to building automation, the company says, but decentralization is not a free-for-all. Regulated industries may require centralized build access for security and risk management, and decentralized automation requires standards for process understanding, tool selection, documentation, and data security. m:…2633

n8n says it is testing its own answer internally, with an internal initiative to demonstrate product capabilities and an automation team transitioning toward a guild model that provides reusable templates, office hours, and documentation standards. The company predicts that firms that decentralize successfully will move faster than those that keep automation centralized, and that automation literacy will become a basic technical skill similar to using the internet. That is a prediction, not a result, but it matches the direction of the week's other launches. m:…2633

Agents that work inside other tools

Two more launches fill in the picture. Rilla Labs says it has launched the Rilla MCP, a read-only connector that brings Rilla's sales conversation data into Claude, ChatGPT, and Cursor, without an API key and without bulk data exports, and that it respects existing role-based access controls. Polyaxon says its platform lets users host an agent controller as a job or service to coordinate machine learning work, with agents that inspect data in a sandbox, launch training and evaluation jobs, read results, and decide which experiment to try next. The controller keeps a proposal-to-execution mapping, and Polyaxon says running inside the platform does not grant the controller additional access by itself. The race to be the default way to build agents is being fought in protocols, services, and hiring models at once, and this week, every front moved. m:…4325 m:…9551