What Happened
Google Labs released Stitch Skills, a standardized plugin architecture that wraps MCP servers for cross-agent portability, hitting #4 on GitHub trending. GitHub published a post-mortem showing how upgrading Copilot’s code review tools actually degraded performance until they adopted Unix-style composable utilities. Microsoft shipped Flint, a declarative JSON-based visualization language that replaces imperative plotting for agent-generated charts. Researchers proposed a multi-agent firewall architecture that sanitizes LLM traffic at the network perimeter. Context.dev (YC S26) launched an API that uses LLM-powered schema extraction to replace brittle CSS selectors for web scraping.
Why It Matters
These releases expose a shared infrastructure problem: agent tooling has outpaced the orchestration layer. Google’s Stitch Skills addresses distribution and versioning across heterogeneous agent platforms. GitHub’s findings challenge the assumption that richer APIs improve agent performance—composition matters more than power. Microsoft’s Flint and Context.dev’s schema-based scraping both narrow the agent’s decision surface by moving complexity into declarative contracts. The privacy firewall research shows enterprises need perimeter controls before LLM traffic leaves the network. Together, these signal a maturation phase where tool boundaries and data contracts become first-class design concerns.
Key Trends
Declarative Contracts Over Imperative Code: Both Flint and Context.dev replace code generation with schema-validated JSON specs. Flint constrains chart generation to data mappings and chart types, offloading rendering to the engine. Context.dev maps HTML to user-defined schemas without CSS selectors. This pattern reduces cognitive load and limits error blast radius by separating what agents decide from how systems execute.
Composition Beats Monolithic Tools: GitHub’s code review post-mortem reveals that richer API calls degraded performance until they adopted Unix-style composable utilities that expose evidence incrementally. The fix wasn’t better tools—it was smaller, chainable primitives that let agents build context step-by-step. This contradicts the prevailing framework assumption that agents need powerful, all-in-one APIs.
Standardization for Cross-Platform Portability: Stitch Skills implements the Agent Skills open standard to package MCP servers as portable plugins across Codex, Claude Code, Cursor, and Gemini CLI. The sparse checkout patterns and three-tier architecture (skill boundary, MCP bridge, marketplace distribution) show how Google thinks about versioning and dependency management at repository scale. This addresses the fragmentation problem as agent platforms proliferate.
Privacy Controls at the Perimeter: The multi-agent firewall architecture intercepts LLM traffic via browser extensions and API proxies before it leaves the network. The multi-agent pipeline detects and redacts PII, credentials, and proprietary code while preserving multi-turn context and tool execution flow. This shifts privacy enforcement from model-level guardrails (which run after data transmission) to network-level interception, critical for enterprises with compliance requirements.