TLDR Dev 2026-11-20
OpenAI’s performance factory 🏭, hardware is software 🔩, PostgreSQL 19 🌱
Together With
Region: Earth is taking over Vegas. Dec 1-3. (Sponsor)
Join Cloudflare in Vegas for a week of content, networking, meetings, and unforgettable experiences, with strategies to take your business to the next level.
- Build It workshops - Guided workshops where you build a project on Cloudflare. Bring a laptop; we'll provide a Cloudflare account if needed.
- Expert Bar - One-on-one conversations with Cloudflare experts.
- Wellness Zone - Take a break with healthy snacks, massages, and quiet rest areas.
Register for free now
How OpenAI's Performance Factory Keeps ChatGPT Fast (15 minute read)
When ChatGPT slows down, one Codex session investigates and a second tries to rule out its explanation before a fix is proposed. In one case, the reviewer reopened an investigation after noticing that the service blamed for failed image downloads had already recovered while the downloads were still failing. Engineers decide what ships, and Codex checks whether performance returns to baseline after deployment.
Stripe Evals for Real-World Coding Agent Performance (13 minute read)
Stripe tested coding agents on one-time payments, subscriptions, and Connect integrations. Agents met 78.1% of required criteria on average, but only 29.4% of runs met every required criterion. All 14 agent setups scored higher with Stripe's plugin.
Distributed, High-Performance Blob Storage on SQLite (12 minute read)
Zerodha built Blobasaur, a Redis-compatible store that shards compressed UI preferences across SQLite files and batches writes per shard. In a memory-backed benchmark with 64 KiB JSON payloads, moving from one shard to 16 raised write throughput from about 7,700 to 25,300 operations per second. Cluster slots are configured statically, with offline resharding and no automatic failover.
Top 10 Things AI Gets Wrong in Web Development (10 minute read)
Coding agents passed 52% of Chrome's web development checks on their own and 91% with Modern Web Guidance installed. Common misses included restoring keyboard focus when a dialog closes, respecting reduced-motion preferences during smooth scrolling, and reserving space for content that loads later.
When Can You Ship on Offline Evals Alone? (10 minute read)
Spotify explores when offline AI evaluations can support a launch decision without an A/B test. In its working example, measurement noise makes the offline eval appear to explain 24% of the variance in online effects when the modeled value is 60%. The proposed rule sends borderline changes to A/B tests and randomly audits some other decisions to detect drift.
Nobody's waiting on code anymore. (Sponsor)
The hard part is context, and knowing what your agents did and why. With Jira, your agents start with the spec and your team's decisions, and every session is tracked. Manage your agent swarm. But keep the receipts.
Learn more at jira.dev
PostgreSQL 19 Released (6 minute read)
PostgreSQL 19 introduces concurrent table repacking to reclaim space while reads and writes continue. It also lets applications wait for a standby to replay a specific write and return an existing row directly when an insert conflicts, simplifying get-or-create queries.
Meet the New Docker, Built to Run on Trust (5 minute read)
Docker unveiled a new brand focused on running AI agents with isolation and oversight. The announcement points to Docker Sandboxes for isolated coding environments and Docker AI Governance for managing agent access and activity.
DataCat (GitHub Repo)
Replit's self-hosted observability stack collects metrics, logs, and distributed traces and provides dashboards and alerts. It accepts OpenTelemetry and Datadog-compatible traffic, with PostgreSQL as the default store and ClickHouse as an option for telemetry.
Become a founding member of TLDR’s AI community (2 minute form)
TLDR is launching a curated community for AI builders to discuss major news and launches. The best links and comments will be featured in TLDR AI.
Apply here to join.
Rebuilding My Voice Assistant Pipeline (6 minute read)
Platypush swaps in microWakeWord for wake-word detection and uses Silero with an adaptive noise floor to tell when someone has finished speaking. In synthetic tests with background chatter and bursts of noise, the old pipeline waited 6 to 10 seconds after the last word, while the new one waited 1 to 1.6 seconds.
Why Chrome Still Uses the Old IP After You Edit Your Hosts File (4 minute read)
Chrome can reuse a connection to the old IP after a hosts file edit because it pools connections by hostname and port. A used socket can stay open for five idle minutes, and flushing Chrome's socket pools forces a fresh lookup but also closes unrelated connections in the same profile.
Coding Harnesses Are Cheap to Swap Now. Model Behavior Isn't. (5 minute read)
Hosted agent APIs and shared formats make harnesses easier to swap, while model choices still change how agents work. With the same repository instructions and tools, Gemini queried a code graph while Claude generally searched from the shell, illustrating why teams should retest behavior after a model swap.
The most important software engineering news in one daily email
Join 480,000 readers for
one daily email