Agentic AI, LLM engineering, and the models behind modern automation — multi-agent systems, LLM evaluation and comparisons, RAG, fine-tuning, AI infrastructure, security, and production AI engineering.
From the blog
Keep Claude agent runs fast and cheap with prompt caching, batching, model tiering, and context discipline — cut cost and latency by half without losing quality.
Use prompt caching, batching, model tiering, and tight context scoping to keep Claude multi-agent runs fast and cheap without losing quality.
Fix the three failure modes that break Claude agents in production: runaway loops, wrong tool selection, and hallucinated arguments — with concrete guards.
Catch loops, wrong tool calls, and hallucinated arguments in Claude multi-agent systems with provenance checks, no-progress detectors, and durable logs.
What to include and leave out of each agent's context in multi-agent Claude systems: orchestrator deltas, need-to-know subagents, compaction, and shared stores.
Design Claude security-agent context right: what to include, what to leave out, how to fence untrusted content, and how to keep reasoning calibrated and cited.
Connect tools and MCP servers to multi-agent Claude systems: per-role auth, self-documenting schemas, typed error contracts, and idempotency for parallel subagents.
Connect Claude security agents to tools via MCP safely: auth, strict schemas, structured error handling, and idempotency so containment never misfires.
Code-level Claude patterns for security agents: typed tools, layered context, evidence ledgers, confidence gates, and deterministic wrappers that hold up.
