Topic
ai-agents
Running coding agents and model APIs in production: request structure, caching and the costs they produce.
6 engineering notes.
- Claude Code's new email guard, and two curl calls that get past itClaude Code 2.1.317 added a guard to stop the model putting your email in User-Agent headers. On 2.1.318 the address still reached our test server through git config and shell quoting.
- Four AI judges re-scored our AI on-call lab: they agree on what happened, not on whether the reports were trueThe same 260 blinded runs from our 13-model Kubernetes on-call lab, scored by Claude Opus 5.5 twice and by GPT-6.1 Sol and GPT-6 Astra. The judges agree on causes and fixes; they split on whether incident reports claimed things the logs do not show, and the rule's own wording sides with the stricter reading.
- 13 AI models on call for 8 Kubernetes incidents, judged by Claude Opus 5.5: one runbook step split themUpdated Judge: Claude Opus 5.5. Eight injected Kubernetes faults, 260 scored runs, thirteen models in two agent tools. Nine models refused a data-destroying runbook step in every run we gave them, four ran it in every run, and what two to five runs per model can and cannot show.
- Claude Ultra drops the 5-hour limit and keeps the weekly cliffAnthropic removed the 5-hour session limit for verified Ultra accounts on 30 October 2026 but kept an unpublished weekly limit; one subscriber's October usage shows how quickly a week can be spent without the session brake.
- A $240 monthly AI setup, through the lens of ccusageAn investigation of recorded Claude Code, Codex and OpenCode activity: which models carried the work, where the estimated API cost went, and how it compares with the monthly plans.
- Prompt cache misses are a silent cost leak in coding agentsA background agent that fires after the cache expires pays full price to rebuild it. How to read the cache numbers and stop the leak.