AI Ops

How prompt caching works on Claude, OpenAI, Gemini and DeepSeek: price multipliers, TTL, a break-even formula, OpenRouter measurements and why cache misses happen.

What context rot is, why LLM answers degrade long before the context window fills, what Chroma, NoLiMa and RULER found, a Claude Haiku 5.5 test and fixes.

When is a Claude Code subagent worth it? Startup cost and result size from the author's runs, what cuts startup context, model choice, and when subagents hurt.

What a token is in an LLM, how many tokens a word or page takes in English and Russian, how to count them before and after a call, and what they cost.

Claude Opus 5.5 vs Sonnet 5.5: pricing, adaptive thinking, the shared tokenizer, and use cases. A decision framework for picking the right model for each task.

Design AI agent memory without a data swamp: session state, durable facts, retrieval, provenance, conflicts, privacy, deletion, and production evaluation.

Secure remote and local MCP servers with OAuth audience binding, least-privilege tools, prompt injection defenses, sandboxing, audit logs, and tests.

A practical production LLM stack: request contracts, model routing, tools, validation, observability, evals, cost controls, fallbacks, and safe rollouts.

A practical AI agent testing workflow for tool policy, grounded answers, execution budgets, model-based checks, human review, and CI release gates.

A deep comparison of context engineering and RAG: when long context replaces retrieval, when you actually need RAG, a decision framework, and the modern 2026 stack.

A complete guide to AI agents: architecture, frameworks (LangGraph, CrewAI, OpenAI Agents SDK, Google ADK, Pydantic AI, MCP), use cases, and a hands-on tutorial.

A 5-component prompt template for building a reusable library. Structure, examples for different tasks, plus organization and versioning of a prompt library.

Reduce LLM spend without hiding quality loss: attribute costs by task, remove waste, set output budgets, test smaller models, use caching safely, and control retries.

An evidence-based tour of KAOS v0.3.0: durable turns, layered memory, MCP tools, governance, behavior evals, portable agent state, and safe defaults.

Design human oversight for AI actions with risk tiers, approval state machines, calibrated routing signals, reviewer UX, queue SLAs, audit logs, and release metrics.

Run prompt experiments on a fixed dataset first, then canary two labeled versions in production. Includes sticky assignment, guardrails, and tested code.

Build a Langfuse prompt workflow in Python: get_prompt(), compile(), version labels, dataset evals, trace linking, canaries, and safe rollback.

Multi-agent architecture patterns for production: Sequential Pipeline, Parallel Fan-Out, Classifier+Router orchestration, task routing, and agent specialization.

Design an LLM judge with narrow rubrics, human calibration, deterministic checks, bias tests, CI gates, production sampling, and safe handling of untrusted outputs.