Ephemeris — the daily dossier
The machinery under the agents.
Today's issue is about the plumbing: a new model to run the agents, a way to deploy anything, and the traces that show you where it all breaks.
Frontier work, now at Sonnet money.
Anthropic shipped Claude Sonnet 5, a model that lands close to Opus 4.8 on coding and agentic work while staying in the Sonnet price tier. It drives a browser and a terminal, and it becomes the default on the Free and Pro plans — so the cheapest way to run an agent just got a lot more capable.
Any Dockerfile is now a Vercel deploy.
Vercel will now run any HTTP server straight from a Dockerfile — Rails, Django, Spring Boot, Go — with autoscaling on Fluid compute. The frontend platform now hosts the whole backend too, so there's no separate container service to wire up alongside it.
The crash was rare. The bug was 18 years old.
OpenAI engineers treated their crashes like an epidemic: mine every core dump at scale, look for the shared symptom. The trail led to two culprits at once — a hardware fault, and a software bug that had been sitting in the stack for eighteen years.
$ analyze coredumps --at-scale --group-by=signature
→ rare, sporadic infra crashes cluster on one pattern
[hardware] a faulty component corrupts memory in flight
[software] a latent bug, ~18 yrs old, does the rest
# neither was visible from a single stack trace
Your requests are already traced. Export them.
Next.js instruments your request path out of the box — you just can't see it anywhere else. Sentry walks through piping those native spans to any OTLP backend with @vercel/otel and a few lines of config, so the traces you're already paying for stop being locked in.
| span | duration | |
|---|---|---|
| GET /checkout | 412 ms | ████████ |
| db.query orders | 188 ms | ████ |
| fetch payments API | 96 ms | ██ |
| render page | 41 ms | █ |
The lab notebook that runs the analysis.
Claude Science is a workbench built for researchers: it wires in the scientific databases and packages they already use, runs the analysis, and returns charts, tables, and paper drafts — with every artifact auditable back to its inputs. It's meant to scale from a single GPU up to a cluster as the question grows.
Nine ways to watch the model.
If you're shipping an LLM feature, "it works in the demo" isn't monitoring. PostHog lines up nine observability tools across the axes that actually decide the choice — tracing, cost tracking, evals, and prompt management — so you can pick on fit instead of logo.
- 01Tracingfollow one request through every model call
- 02Cost trackingtokens and dollars, per feature
- 03Evalsdid the output get better or worse?
- 04Prompt mgmtversion and roll back the prompt
That's today.
Six picks from the feeds on Wednesday, 01 July 2026 — a tools-and-infrastructure day.