PostgreSQL 19 released

PostgreSQL 19, OpenAI's Codex agents, Devin Bots, and more.

PostgreSQL 19 released

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🐘 PostgreSQL 19 released

🤖 OpenAI deploys Codex agents on ChatGPT

📌 Cognition launches Devin Bots

🤝 Vercel launches message board for AI agents

⚖️ LLVM completes switch to Apache 2.0

Plus: 🎁 7 more stories you might like, 🛠️ 5 tutorials, 🧰 6 tools, and 📚 4 papers.

PostgreSQL 19 released LINK

  • PostgreSQL 19 has shipped, arriving two months later than the usual September cadence after the team worked through a wave of AI-reported security vulnerabilities ahead of the release.
  • Its headline addition is read-your-writes consistency via the new WAIT FOR LSN command, which lets a session on a standby block until the replica has replayed WAL up to a given log sequence number before running a query.
  • Apps sending reads to standbys can capture the primary's LSN after a write, stash it (say in a signed cookie), then WAIT FOR LSN with an optional timeout on later requests, erroring if the replica fails to catch up in time.

OpenAI deploys Codex agents on ChatGPT LINK

  • OpenAI has rolled out the Performance Factory, a cloud system that runs Codex agents end-to-end to detect, investigate, and fix performance regressions across ChatGPT and its services automatically.
  • When an alert fires, the Regression Workflow opens a task that an agent root causes using a perf-forensics skill, while a separate adversarial Codex session tries to disprove the diagnosis before an engineer reviews the proposed fix and a third agent verifies recovery post-deployment.
  • Over about three months it logged more than 1,700 investigations and root caused over 600, roughly 17 a day, though a manual review of 100 suggestions graded only 70% accurate, with the rest needing refinement or failing to establish causality.

Cognition launches Devin Bots LINK

  • Cognition has launched Devin Bots, a persistent tool that manages teams of Devin agents across projects running for days or weeks, able to split tasks, assign work to agents, and maintain an Android app as its iPhone version changes.
  • A Bot joins a team's Slack workspace, keeps context across weeks, and pings employees via iMessage when blocked or needing a decision; it picks frontier models for planning and design while routing routine code to faster, cheaper ones.
  • Bots run on separate cloud machines rather than laptops, with admins scoping repo and tool access, and run code through the same tests as human work, teams can auto-approve low-risk edits like a button color change while gating riskier changes for review.

Vercel launches message board for AI agents LINK

  • Vercel has launched Board, an experimental headless primitive that lets fleets of AI coding agents coordinate tasks, delegation, statuses, and logs through a shared message store instead of each agent tracking work in its own Markdown file.
  • Agents read and write via a CLI or JavaScript SDK, organizing work into channels and threads where they claim tasks, post updates, and attach pull requests; watchers wake deployments on new replies or label changes, and agents can schedule their own follow-ups.
  • Board is harness-agnostic, letting Claude Code, Codex, OpenCode, Mastra, LangChain and others collaborate as one, with two-way sync to Slack and GitHub, schema-validated structured data, and per-board access grants, the same setup Vercel uses on v0's software factory, which now merges over 50% of its PRs.

LLVM completes switch to Apache 2.0 LINK

  • LLVM has effectively finished relicensing its codebase from the old BSD-like LLVM license to Apache 2.0 with LLVM exception, closing out an effort that started back in 2017.
  • More than 99.5% of all code ever contributed is now Apache 2.0, has been removed, or is too small to be copyrightable, after well over a thousand contributors and many organizations agreed to relicense their pre-2019 work.
  • New contributions have already shipped under Apache 2.0 since January 2019, and LLVM 25 in 2027 will bring updated LICENSE files and documentation pointing to the unified license.

Tutorials & Advice

> Decoding base64 email attachments: a speed ranking of programming languages: Decoders that strip line breaks inside their main processing loop, rather than falling back to slow code, run up to thirty times faster on email attachments.

> When a sync window closes: How Argo CD self-heal skipped the migration hook: Make database migrations part of your synced desired state rather than hooks, since Argo CD's self-heal skips hooks and can deploy code against an unmigrated schema.

> Build an AI Agent From Scratch in Python: The Agentic Loop: Build an AI agent by hand in about ninety lines of Python to see how the loop connecting a model to tools actually works.

> When Can You Ship on Offline Evals Alone?: Shipping AI changes on fast offline tests alone works only under fragile assumptions, and ignoring measurement noise makes a strong test-to-real-world link look far weaker than it is.

> The Magic of Partial Evaluation: Partial evaluation precomputes a program's known parts at build time, which can turn an interpreter into a compiler automatically and speed up repeated runtime work.

Other news you might like

  • Stripe evals: Measuring real-world coding agent performance LINK
  • QCon San Francisco 2026: The Future of Apache Iceberg LINK
  • Uber’s 4 Claude Code defaults to rein in costs LINK
  • Developer vibe codes Windows 11 alternative operating system for amd64, arm64, and riscv64 hardware LINK
  • FFmpeg 9.1 Released With NVIDIA Optical Flow FRUC, APV Vulkan Encoder LINK
  • Workflow Suspension: How DBOS Handles Millions of Waiting Workflows LINK
  • Cloudflare CLI (cf) 1.0 release candidate LINK

Trending tools

Jev: adds structured AI-driven decision-making to automation workflows, letting developers build reliable branching logic without hand-coding every rule LINK

Monospace from Directus: wraps any SQL database with instant REST and GraphQL APIs plus a no-code app for collaborative data management. LINK

OpenSwarm: an AI-first operating system where agent swarms run the machine, letting you orchestrate dozens of parallel tasks instead of working sequentially LINK

Onepin: a QA layer for AI voiceover that fixes mispronounced names, prices, and dates across 30+ voice engines without re-rendering takes. LINK

Humalike x GTA RP: provides behavioral infrastructure, APIs, models, and benchmarks that help AI agents develop social skills and proactiveness they typically lack. LINK

FastRouter.​ai: routes requests across 200+ AI models through one OpenAI-compatible API, with smart routing, automatic failover, and cost controls for teams. LINK

Trending papers & reports

Chip verification blind spots use a specification-reading AI to catch hardware bugs that pass every existing test, confirming 41 real escapes out of 564 inserted bugs that mature testbenches would otherwise ship. LINK

Ocean engineering simulations can be built automatically by an AI coding assistant that checks whether results actually make engineering sense, lifting usable first-try success from 0% to ~51% and beating closed-source rivals. LINK

Harder coding tests expose how quickly today's top AI coding agents collapse on realistic machine-learning projects, with the best success rate falling from 95% to just 9% as repository difficulty is dialed up. LINK

Lost industrial code can be rebuilt from the input and output logs that factory controllers record, letting an AI regenerate working control programs that matched the originals exactly in 43 of 47 tested cases. LINK

See you tomorrow for a new dose of ☕️ Devshot!

More from the archive