PLUMBING.
Today's issue is about the layer nobody demos — the logs, the name resolution, the eval math, the walls between agents — and why it decides whether the demos survive production.
Prove the cheaper model is actually cheaper.
Open weights are surging and every pipeline is a migration candidate — but a single eval run flatters whichever model got lucky. oss-migration-eval, a Claude Code skill from Bayram Annakov, decides the swap with statistics: human-validated golden datasets seeded with trap cases, provider-pinned sweeps across several seeds, deterministic grading, and an optional blind panel for subjective work. It reports cost saved against accuracy lost, with confidence intervals instead of vibes.
- ~$1
- a full 9-model × 8-case × 5-seed sweep on OpenRouter
- 3–5
- seeds per model, so one lucky sample can't crown a winner
- CI
- savings vs accuracy loss reported with confidence intervals
A log is an event, not a sentence.
Sentry's field guide treats a good log as a small, queryable event: a stable name in domain.action form, dynamic values pushed into flat dot-notation attributes, and every value kept to a primitive. Do that and each field becomes something you can filter, group, and alert on at 3am — no regex archaeology across free-text messages.
One resolver for both sides of the network.
Cloudflare Internal DNS is generally available: authoritative internal zones and recursive public resolution now share a single control plane on Cloudflare's edge. A Zero Trust network can resolve private hostnames — split-horizon views, internal-only records — without standing up and hardening its own DNS servers, and the same dashboard answers public queries.
Give your agents separate rooms.
The point of a multi-agent council is disagreement. But hand the members shared file access or a shared context and they read each other's reasoning first — the diversity you built the panel for collapses into groupthink. Bayram Annakov's /council skill fixes it in the architecture: each persona evaluates inside its own sub-agent, and the evaluator runs in a separate context window that never sees the generator's chain of thought.
Mind the plumbing.
Four picks from the layer under the demos: a statistical way to size a model migration, a grammar for logs you can query, one resolver for private and public names, and the wall that keeps an agent council honest.
Sources today: Sentry · Cloudflare · Telegram @ProductsAndStartups. Rubric: tools to adopt this week, dev & agent infrastructure, security and research with a practical kernel.
Ephemeris · Issue 093 · 23 July 2026 · Zürich