Daily Technical Brief: Agent Infrastructure Matures Around Legacy Integration

What Happened

The past 24 hours reveal a sharp focus on bridging legacy systems and modern agent frameworks. Virtual MCP servers emerged as a config-driven pattern for exposing internal APIs without custom code. Browser Use hit 120K GitHub stars by separating agent reasoning from browser primitives. Pomerado launched self-healing web scrapers as MCP servers. Hypercubic’s Hopper demonstrated mainframe-to-LLM bridges for COBOL and JCL systems. Meanwhile, a KDD 2026 study compared three agent evaluation paradigms, finding no single framework catches all failure modes.

Why It Matters

Production agent deployments now depend on integration plumbing, not just LLM quality. Virtual MCP servers reduce the cost of exposing internal services from weeks of TypeScript to hours of YAML. Self-healing scrapers address the contract stability problem that breaks browser automation in production. Mainframe bridges unlock decades of business logic without rewrites. These patterns share a common architecture: adapters that translate legacy interfaces into agent-callable tools while preserving existing security models. The evaluation study matters because it quantifies what practitioners already know—trace-based, outcome-based, and LLM-judge frameworks each catch different failure classes, forcing teams to run multiple eval pipelines.

Key Trends

MCP as the universal adapter protocol. Four of five articles describe systems that expose non-agent-native interfaces (OpenAPI specs, websites, mainframes, browsers) as MCP servers. The pattern: read an interface definition, generate tool schemas, proxy requests. Virtual MCP servers treat OpenAPI/GraphQL schemas as tool definitions. Pomerado converts DOM selectors into deterministic extraction APIs. Browser Use wraps Playwright in MCP tool contracts. This convergence suggests MCP is becoming the de facto agent tool protocol, similar to how OpenAPI standardized REST documentation.

Self-healing as a requirement, not a feature. Both Pomerado and Browser Use emphasize automatic recovery from interface changes. Pomerado detects DOM mutations and rewrites selectors. Browser Use offers stealth modes to bypass bot detection. The technical insight: agents need contract stability even when underlying systems change. This shifts complexity from “write a scraper” to “maintain a scraper that rewrites itself.”

Legacy system integration drives real-world adoption. Hypercubic’s mainframe bridge highlights the gap between agent frameworks and production infrastructure. Mainframes have no JSON, no REST, no streaming—just TN3270 screens and COBOL. The solution involves TN3270 protocol adapters, RACF security passthrough, and JCL job submission APIs. This is harder than exposing a modern REST API but unlocks systems processing 70% of global transactions.

Evaluation remains fragmented. The KDD study tested trace-based, outcome-based, and LLM-judge frameworks on analytics agents. Trace-based evaluation caught subtle functional errors. Outcome-based was simpler but noisier. LLM-judges surfaced issues the other two missed. No single framework covered functional correctness and security risks. Implication: production teams need multi-paradigm eval pipelines, not a single golden metric.

Hot-swappable modules for agent iteration. Simon Willison’s Beep Locator demonstrates WebSocket-based module injection to let agents iterate on browser-based listeners without permission prompts. The pattern: Python orchestrator deploys JavaScript to browsers, coordinates uploads, runs inference. This architecture separates agent reasoning (Python) from execution environment (browser) and enables rapid iteration on live systems.