> blog
Field notes from building skills, plugins, and agents in 2026. MCP, Claude 4.6, RAG, evals, and the production patterns we wish we'd known earlier.
MCP Mastery — A Field Guide for 2026
The Model Context Protocol is the integration standard AI agents will use for the next decade. Here's what you need to know to build, ship, and secure MCP servers in 2026.
mcpmodel-context-protocolai-agentsclaudeThe Claude 4.6 features you should ship today
1M context, extended thinking, the memory tool, code execution, 1-hour prompt caching, computer use. A tour of what's production-ready in Claude 4.6 and how to wire each one into a real app.
claudeclaude-4-6anthropic-apiai-engineeringPrompt caching at 1h TTL: cutting LLM spend by 80%
Most teams underuse prompt caching. Here's the exact pattern for placing cache breakpoints, the mistakes that silently disable caching, and how we cut one workload's bill by 80% with a single line of config.
prompt-cachingcost-optimizationclaudellm-opsEval engineering: the regression gate no one built
You wouldn't ship a web app without tests. You probably ship AI features without evals. Here's the minimum viable eval engineering setup that catches regressions before users do.
evalsllm-as-judgeregression-testingai-engineeringSub-agent delegation patterns that actually work
Sub-agents aren't a magic concurrency trick. Used right, they parallelize research and protect context. Used wrong, they're expensive token thrashing. Here's how to tell the difference.
agentsclaude-agent-sdksub-agentsai-architectureWriting SKILL.md: 4 rules for reliable activation
Agent skills are easy to write and hard to write well. Four rules for SKILL.md frontmatter and body that make the difference between a skill that fires and one that gathers dust.
skillsskill-authoringclaude-codeagents