An agent that takes 47 steps, calls 12 tools, and costs $0.23 to run is a black box without proper observability. Tracing, metrics, and cost attribution for agentic systems — what to instrument and how to make the data useful.
How do you know your agent is getting better rather than worse? Evaluation for agentic systems is harder than for single-turn LLM calls — tasks span multiple steps, outputs are open-ended, and human review doesn't scale. The evaluation architecture that makes continuous improvement possible.
The EU AI Act's main obligations entered full enforcement in August 2026. This is what the Act actually requires of engineering teams building AI systems — not the legal overview, but the specific technical and documentation work that compliance requires.
Autonomous agents can take actions that are hard or impossible to reverse. The engineering principle of blast radius control — limiting what can go wrong when an agent makes a mistake — and the specific patterns that implement it.
Prompt injection is the attack surface that didn't exist before LLM-based agents. An agent that reads external content — web pages, emails, documents, database records — can be manipulated by that content. The attack taxonomy and the defences that actually work.
The Cloud Security Alliance found security flaws in 62% of AI-generated code samples they tested. What types of flaws, why AI generates them, and the review and tooling practices that catch them before they reach production.
Single tool calls are easy. Production agents orchestrate dozens of tools across parallel branches, handle failures gracefully, version tool interfaces, and manage cost across thousands of calls per session. The patterns that make tool use reliable at scale.
Long-term memory is what separates agents that get smarter over time from agents that start from scratch on every session. The four memory patterns, how to implement each, and the engineering decisions that make agent memory practical.
The 2026 pattern for cost-efficient production agents: small language models handle simple tasks cheaply, large models handle complex tasks well. How to design the routing layer, what to measure, and what the cost-quality tradeoffs actually look like.
Reasoning models think before responding — generating extended chain-of-thought that improves accuracy on hard problems at the cost of latency and tokens. When the thinking investment pays off, and when it doesn't.