Ephemeris.
Today's issue is about the gap between what agents can do and what we let them see — production context, KPI gaming, code that survives review, and the talent the world is trying to keep at home.
Engineering · Observability
The bottleneck is not intelligence.
Sergiy Dybskiy's argument from Sentry's engineering blog: self-healing agents stall not because they are dim, but because they have no idea what actually broke in production. Hand them traces, breadcrumbs, and the failing user's session — and the loop closes. Without that, they guess.
# agent fix attempt — no production context $ claude --task "fix the timeout on /checkout" → scanning repository... → hypothesis: rate limit middleware? → patch attempt 1, 2, 3 ... FAIL · cannot reproduce locally # same task, with Sentry MCP wired in $ claude --task "fix the timeout on /checkout" → pulling trace 9b2a — 12.4s on getCart() → N+1 on user.addresses, 412 rows → patch: eager-load addresses PASS · p95 down 7.8s → 240ms
Essay · Long view
Reckoning with the future.
Clark's longest essay in months arrives wrapped, as he often does, in a short story — this one about a singularity that turns out to be neither apocalyptic nor utopian, just busy. The story is bait. The argument behind it is that the people who built today's frontier models are running out of time to write the institutions that ought to hold them.
He names the awkward middle the field is in. Capability climbs faster than anyone, including the labs, can metabolise. Regulators reach for the closest analogue — pharmaceuticals one week, weapons the next — and get neither right. Founders ship and apologise. Researchers publish papers their employers privately wish they hadn't.
The fictional frame lets him do something the essays alone could not: imagine a future where the choices were good ones, and trace back what was true in the years before. The list is short and uncomfortable. Pay close attention to it. Today, while there is still time, is when the small decisions count.
A1 · Q1 · OBSERVED
When the KPI is the token.
Amazon set a target — eighty percent of developers using AI tools weekly — and stood up an internal leaderboard tracking token spend per team. Employees responded the way every reader of Goodhart already knows they would: they burned tokens on meaningless tasks to look compliant, and the metric stopped meaning anything within a quarter.
| A — team | B — tokens / dev | C — useful tasks | D — status | |
|---|---|---|---|---|
| 1 | retail-promos | 1,184,000 | 12 % | gaming |
| 2 | fulfilment-ops | 902,400 | 18 % | gaming |
| 3 | aws-billing | 612,800 | 41 % | mixed |
| 4 | video-search | 248,300 | 71 % | on target |
| 5 | kindle-content | 96,400 | 83 % | on target |
| 6 | identity-core | 14,200 | 91 % | below quota |
Research · Coding agents
Forty-four percent survives.
SemiAnalysis dissects a Stanford study of six thousand real coding-agent sessions and frames it against the economics of the assistant market. The number to remember is forty-four — the percentage of AI-generated code that survives review and reaches the main branch. Pure vibe-coding fares worse still.
44%
Code that survived to commit
Across 6,000 sessions. The other fifty-six was deleted, rewritten, or rolled back by the reviewer the same day.
9×
Vulnerabilities, vibe vs human
Pure vibe-coded paths shipped 0.76 vulnerabilities per kLOC; human-led paths shipped 0.08. Verification, not generation, is the new bottleneck.
Geopolitics · Talent
Beijing tightens the passport.
Travel restrictions that once applied only to state officials have been quietly extended to senior AI engineers at private firms — Alibaba, DeepSeek, and the second tier behind them. Bloomberg's reporting points to a policy that treats model-builders as strategic assets, on the same footing as nuclear and semiconductor staff.
ScopeSenior research staff, top-tier labs
MechanismPassport surrender, exit-permit review
PrecedentAlready used for SOE, defence, semi staff
Named firmsAlibaba, DeepSeek, others unconfirmed
TriggerRecent talent-poaching abroad, IPO season
ImplicationHire from outside, or not at all
Dev tools · Patterns
A working set of fifty patterns.
OpenAI quietly opened a use-case library for Codex — not a prompt zoo, but a filterable lattice of real workflows by category, role, and task type. Inbox triage, design-to-code, GitHub review, onboarding flows: fifty-plus entries with implementation notes. The interesting move is the taxonomy itself.
01 · ENG
iOS / macOS app scaffolds with native toolchain
02 · QA
Bug-bash triage from a failing trace
03 · DATA
Ad-hoc analysis on a CSV with checked assumptions
04 · DESIGN
Figma frame to React component, with tokens
05 · OPS
Onboarding plan generated from the repo README
06 · GH
Review a PR against the team's style guide
07 · INBOX
Email triage and reply-drafting on the calendar
08 · MEET
Pre-read assembly from CRM and shared docs
That's today.
Six files, six sources, one theme: the distance between what models can do and the production context they need to do it without supervision.
Sentry Engineering · Import AI · Techspot · SemiAnalysis · Bloomberg · OpenAI Developers · with via from @ProductsAndStartups, @seeallochnaya · Rubric: AI tools you could adopt this week; dev tools and agentic coding; research with a practical kernel; anything immediately actionable for a senior engineer or founder.