Nimble | AI Engineers Radar

MCP Ecosystem Expansions

• 🔥 Citrix MCP Gateway: Citrix announced NetScaler MCP Gateway with OAuth 2.1 support, traffic controls, and unified...

New Benchmarks for Agentic Retrieval

• 🔥 AgenticSTS: Introduces a bounded-memory testbed using Slay the Spire 2 with 298 trajectories for...

Coding agent harnesses are maturing into production systems that combine dedicated IDEs, orchestration layers, and deterministic validation.

• ZCode...

• MCPUseMetric in DeepEval now scores how effectively agents call MCP primitives and arguments via LLM-as-judge.
• DiscoBench introduces 211 samples...

HOLA augments linear attention's compressive recurrent state with a bounded exact KV cache that stores only high-residual tokens, recovering...

Three developments outline a layered governance approach for agents:

• Context provenance via ContextNest delivers verifiable knowledge vaults with...

• No more copy-paste: Sentry MCP with Cursor skips manual error logs and stack traces for RAG debugging
• Production context: MCP investigates real...

A2UI v0.9 lets agents emit structured intent instead of arbitrary UI code, keeping rendering inside trusted app components.

Key production challenges...

MCP Ecosystem & Browser Tooling

• 🔥 Apple Safari MCP Server: Apple embedded a Model Context Protocol server in Safari Technology Preview enabling...

Browser makers are converging on native agent tooling, but the approaches differ sharply in where control lives.

• MCP servers (Safari) deliver...

• MotherDuck Flights enables agents to build, run, and manage Python ingest pipelines via MCP, delivering source-to-analytics in one session with...

JadePuffer's LLM-driven ransomware exploited exposed Langflow and shared credentials to encrypt databases and drop schemas in minutes.

• Model choice...

SkillWeaver's decompose-retrieve-compose pipeline with SAD feedback and DAG execution graphs tackles the core bottleneck of routing across massive...

In production RAG, question parsing and retrieval evaluation are silent failure points that cascade into bad answers.

• Article 1 shows retrieval...

Treating full execution traces as evaluation data shifts agent assessment from pass/fail outcomes to path-level attribution and debugging.

• Harness...

Memory is becoming essential for long-term LLM agents, yet retrieved memories frequently trigger sycophancy by causing agents to over-align with users...

MCP Ecosystem Expansions

• 🔥 Safari MCP Server: Apple released a Safari MCP server in Technology Preview 247 that lets coding agents access page...

Databricks productionizes agents via a tight baseline-metrics-monitor-refine loop powered by MLflow.

• Create small, user-informed eval sets and log...

OpenGov's OG Assist demonstrates production-grade agent deployment for government workflows:

• Architecture pivot: Migrated from LangGraph to a custom...

First-party MCP servers are now shipping in production across five verticals, letting AI agents tap live data without custom infra.

• X hosts...