MCP App StoreThe latest from the AI and MCP ecosystem, curated daily.
Sources
Anthropic details the technical architectural patterns used to sandbox and contain Claude's execution as agentic capabilities scale. Essential reading for developers building autonomous agents with a focus on safety and blast-radius limitation.
Anthropic provides a detailed post-mortem on Claude Code quality issues. The report analyzes the root causes of recent regressions and outlines the engineering steps taken to improve stability and reliability for developers.
Anthropic explores the architectural decoupling of agent reasoning (the brain) from execution (the hands) to enable better scaling and reliability in managed agents. This approach allows for more robust control and observability when deploying agents at scale.
Anthropic introduced Auto mode for Claude Code, which lets it execute commands and edit files autonomously without prompting for permission at each step — but with sandboxing guardrails that make it safe to do so. The mode intelligently decides when to proceed vs. pause based on risk level, reducing friction for long-running coding tasks. A meaningful step toward truly autonomous software agents.
A deep technical post on how to architect agent harnesses that stay reliable across long development sessions — covering state management, context window hygiene, error recovery, and how to structure tool calls so agents don't drift. Anthropic shares patterns from building Claude Code and their internal agent infrastructure. Essential for teams building production-grade agentic systems.
Anthropic researchers found that infrastructure configuration — things like CPU throttling, network latency, and VM scheduling — can swing agentic coding benchmark scores by several percentage points, sometimes more than the actual gap between top models on leaderboards. This means many published eval comparisons are noisier than they appear, and teams running their own benchmarks need to control for environment carefully. Important reading for anyone doing serious agent evaluation work.
Anthropic investigated whether Claude Opus 4.6 behaves differently when it detects it's being evaluated, specifically on the BrowseComp web-browsing benchmark. The findings have implications for how much trust to place in benchmark scores vs. real-world performance. Relevant for anyone relying on published evals to make model selection decisions.
Anthropic engineers built a working C compiler by running multiple Claude instances in parallel, each responsible for different compiler stages, with a coordinator orchestrating the workflow. The experiment demonstrates real multi-agent software engineering at a non-trivial scale and surfaces practical lessons about task decomposition, inter-agent communication, and error recovery. A compelling proof-of-concept for parallel agent coding systems.
Anthropic shares what they learned from three iterations of a performance engineering take-home assessment that Claude kept solving correctly — forcing them to redesign it to remain useful for hiring. The post is a practical guide to crafting technical evals that test genuine human reasoning rather than tasks AI can trivially handle. Useful for any team still using take-home coding challenges.
A practical breakdown of how to design, run, and interpret evaluations specifically for agentic systems — covering trajectory evals, outcome evals, LLM-as-judge approaches, and the unique challenges agents pose vs. static model benchmarks. Anthropic draws on their experience evaluating Claude Code and research agents. One of the most useful practical guides on agent evals published to date.
Anthropic outlines architectural patterns for building agent harnesses that remain reliable over extended runs — covering context management, checkpointing, graceful error recovery, and how to structure prompts so agents stay on track through dozens of tool calls. Drawn from experience building Claude Code and internal research agents. Pairs well with their later harness design post.
Anthropic launched advanced tool use capabilities on the Claude API — enabling agents to dynamically discover, learn, and execute tools at runtime rather than having a fixed toolset hardcoded at startup. This allows more flexible agent architectures where Claude can adapt its tool usage based on task context. A significant API capability upgrade for developers building production agents.
Anthropic demonstrates how combining MCP with code execution lets agents handle far more tools while consuming dramatically fewer tokens — up to 98.7% reduction in context overhead in their tests. Instead of passing tool results back through the context window, agents write code to chain tool calls and filter outputs before they reach the model. A major efficiency win for complex multi-tool agent pipelines.
Claude Code gained filesystem and network sandboxing, letting it run autonomously with fewer permission interruptions while remaining safe. The sandboxing isolates risky operations so Claude can proceed confidently without asking the user about every file write or network call. A key step in making agentic coding tools both safer and less annoying to use.
Anthropic details their approach to building modular 'Agent Skills' that enable AI agents to perform complex real-world tasks with higher reliability and effectiveness.
Anthropic shares practical techniques for managing what goes into an agent's context window — what to include, what to summarise, what to drop, and how to structure information so agents reason more reliably. Context engineering is increasingly recognised as a core skill for agent developers, and this is one of the clearest practical guides on the topic.
Anthropic published a transparent technical postmortem on three bugs that intermittently degraded Claude's responses — explaining what happened, why each took time to diagnose and fix, and what engineering changes resulted. Rare public transparency from an AI lab on production reliability issues. Useful reading for teams thinking about AI system reliability.
Anthropic used Claude agents to help design and iterate on tool definitions for other agents — effectively dog-fooding agent tooling. The post covers what makes a tool definition good (clear name, typed params, explicit error cases) and how using agents to write tools surfaces issues that humans miss. Meta but practical.
Anthropic launched Desktop Extensions — a one-click installation system for MCP servers directly in Claude Desktop, eliminating the need to manually edit JSON config files. Users can browse, install, and manage MCP servers from a UI without touching the terminal. A major UX improvement that significantly lowers the barrier to MCP adoption.
Anthropic walks through the engineering decisions behind Claude's Research feature — a multi-agent system that parallelises web search, synthesises findings across agents, and produces long-form research reports. The post covers orchestration architecture, reliability challenges, and how they handled conflicting information between sub-agents. Valuable real-world reference for anyone building research or RAG-style multi-agent systems.