AI engineering for production
Hands-on articles for engineers building with AI: designing AI systems, working with coding agents, running AI infrastructure and keeping AI costs under control.
CVE-2026-110264 lets an unauthenticated attacker run code on self-managed GitLab, in every release back to 7.0. Which version to move to, why the backports don’t close the whole release, and how to upgrade without a second incident.
DeepSeek’s Search and Fetch APIs cost $0.75 per 1,000 calls against $10 on Claude and OpenAI. For most agents the call fee is the smaller half of the bill: the result tokens are the rest. How to price both and cap them.
Hitting Claude Code’s weekly limit by Wednesday? /usage now shows where the allowance goes. Subagents running on your main model, long sessions and idle loops are the usual suspects, and most fixes are a setting.
OpenAI’s Decisions API returns probabilities instead of a generated JSON label, and bills input tokens only. When that number changes how you route, flag and abstain, and when Structured Outputs or function calling is still the right call.
Mistral Large 4 lists at $1.36 in and $4.18 out per million tokens, far below Claude Opus 5.5 and GPT-6 Astra. Here’s why the launch price, cached input and cost per finished task shrink that gap, and how to measure it on your own traffic.
Small models now cost a fortieth of the top tier, so routing easy requests to them is worth doing. How a cascade decides what escalates, where its hidden costs are (cold caches, long-prompt pricing, wrong answers that pass), and when one model is simpler.
Per-call LLM logs can’t explain why an agent took forty steps or cost ten times the usual. What an agent trace should capture, how to keep one trace across agents and pauses, what not to record, and how to sample.
A long agent run that crashes either starts over, repeating model calls and side effects, or is lost. How durable execution records and replays steps, where step boundaries go, and why idempotency is still your job.
Sandboxing each agent doesn’t isolate them from each other. Injection spreads through the memory, documents and messages they share. Where it travels, why per-agent safety doesn’t compose, and the design that contains it.
Ollama has no authentication for local requests, and vLLM’s --api-key covers only some paths. What an exposed inference server gives away, and the setup that closes it: loopback binding, an allowlisting proxy and closed ports.
A system prompt lowers how often an agent misbehaves; it can’t stop a prompt injection. Where controls actually hold: authorization in every tool, limits once untrusted content is in context, and checks before consequential actions.
Green CI only proves the branch builds and passes the tests it ships with. Read the task, the file list and the test diff before the implementation, and run the new tests against the old code.
