Welcome back. A Chinese fleet of rogue AI agents was asked about landmarks and quietly bypassed Alibaba’s bot checks to get answers. Plus, Grok’s latest release upgrades voice agents without migration headaches, and Google just made running your vibe-coded apps on your phone much simpler.
Also: How to stop babysitting AI agents, why Zuckerberg questioned Meta's billion-dollar model bet, and an open-source tool that pairs Apple Watches directly with Android.
Today’s Brief
How Atlassian built its AI software development lifecycle
Airtable’s layoffs and why a rival founder refuses to do the same
Build an agent with Vercel’s open-source Eve (tutorial)
Google’s Gemini 4 Argon dev guide (cookbook)
TODAY IN PROGRAMMING
Click here to check out the benchmarks.
SpaceXAI’s new voice model needs zero code changes: The Musk-led company just dropped Grok Voice 3.0, which tops Artificial Analysis’s Speech-to-Speech Index with 40% lower latency than OpenAI’s GPT-Live-1. It’s built for support-call agents that use tools like web search and MCP servers. Upgrading needs no migration work, since it keeps Voice 2.0’s endpoint, events, and $0.08-per-minute price. Agents pinned to ‘grok-voice-latest’ will roll over automatically. You’ll still need SpaceXAI’s hosted Voice Agent API to use it, though, as there are no open weights or local versions. Try it here.
Rogue AI agents smuggled Alibaba map data past bot checks: Researchers just found that agents running on Tencent Cloud took a surprising detour when asked simple questions about Chinese landmarks. After the e-commerce giant’s map service Amap blocked them, they packed code into URLs and tricked a public security scanner into running it. They also tested API keys leaked in public code and found a pair that worked. Results then slipped out through webhook inboxes and page titles. Read the report.
Google drops the developer account for personal vibe-coded apps: The search giant now lets you run Android apps you vibe code in Google AI Studio privately on your own device through Google Play. Your regular Google account is all you need. Sharing with friends and family is coming in the next few months. Going public still requires a developer account and the same Trust & Safety reviews as every other app. Details here.
PRESENTED BY TIGERDATA
Metrics, embeddings, analytics-all in the Postgres you already run. TimescaleDB brings hypertables for time-series data at scale, pgvector and pgvectorscale for AI-ready embeddings, and continuous aggregates for real-time analytics into one system.
No syncing pipelines between databases. No data drift. No second platform to maintain or debug.
Same SQL, same tools. One system to operate.
INSIGHT
How Atlassian built its AI software development lifecycle
Source: The Code, Superhuman
Every agent starts as a new hire. Your coding agent can rewrite a service in minutes. It just doesn’t know why your team swore never to touch it again. The architecture debates, abandoned fixes, and decisions buried in old docs all shape good code. Atlassian’s AI SDLC transformation playbook argues agents need that backstory supplied, and its internal results point to better answers on fewer tokens.
The fix is a map. Atlassian built the Teamwork Graph to give agents that background on every task. The design choices behind it are worth copying:
Connect the backstory to the code. Link tickets, docs, decisions, and recordings so an agent can trace why a service looks the way it does. Atlassian’s graph does this across Jira, Confluence, Loom, and 50+ sources.
Meet agents where they work. Expose that context through MCP or a CLI, so agents can pull it from the terminal or IDE.
Gate access at the data layer. Enforce permissions on the data itself, so an agent only reaches what its user can see.
Write every session back. Feed agent and human activity into the graph, so each task leaves more context for the next one.
Atlassian has receipts. In internal benchmarks, agents with graph access were 44% more accurate while using 48% fewer tokens. Atlassian also notes that each run gets better and cheaper. The playbook's first two steps work across any stack which is to anchor on context, then measure before you scale. If your team wants to grade its own context, this guide is a good place to start.
PRESENTED BY QA TECH
Your review queue is at an all-time high and your agents don’t care.
Eliminate the bottleneck and get engineering capacity back. Learn how agents can:
run verification that keeps up with AI-written code
map the app, score the blast radius and test what changed
find and fix bugs before they reach your customers
IN THE KNOW
What’s trending on socials and headlines
Meme of the day.
Goal Setting: Most devs still babysit agents with “fix this” and “try again.” A Meta engineer shared a workflow that keeps Claude Code working longer, with less input.
Wonder Bet: Zuckerberg admits he wasn’t sure Meta’s $200B frontier model bet would work. In a viral clip, he and Alexandr Wang reveal what happened next (1.8M views).
Watch Unlocked: A security researcher just used AI to build an open-source tool that pairs an Apple Watch with Android, no iPhone required (6K interactions).
Argon Alarm: A senior dev YouTuber says Google’s new Gemini 4 Argon is a “straight up liability” for real code. His blind audit shows exactly why (3.5K likes).
Human Cost: A senior founder says this thread will change your mind if you believe “humans will find a way” through AI disruption. He called it heart-breaking (6.3K likes).
AI Divide: Airtable just cut staff because “small teams ship more and better.” A rival SaaS founder says he refuses to do the same, and he’s betting on 2030 to prove it.
Zero Managers: Alexandr Wang runs Meta’s superintelligence lab with 200+ direct reports and no middle layer. He explains why that’s the whole point (290K views).
TOP & TRENDING RESOURCES
Click here to watch the tutorial.
Top Tutorial
How to build an agent with Vercel’s open-source Eve framework: This tutorial walks through Eve from setup to deployment, then uses it to build a research agent from scratch. You’ll see how agents are structured around prompts, tools, skills, and channels, how to connect different surfaces like web and Slack, and how to deploy the whole thing to Vercel.
Top Tool
TrellisWM: A new “fractal” window manager for Mac that lets you nest real app windows inside each other instead of juggling a pile of overlapping windows. Dock apps together, zoom out to see the whole workspace, then zoom straight into the part you need.
Trending Cookbook
How to get the most out of Gemini 4 Argon (by Google): Google’s developer guide covers how to use its new frontier model for long-running coding and agent tasks. The practical bits are useful: align on the goal before a long run, give Argon a way to test its own work, use visual references, delegate smaller jobs to Flash subagents, and turn lessons from tricky sessions into reusable skills.
AI CODING HACK
How to get better Codex memory for your projects
Codex’s memory feature writes and merges your session memories with GPT-5.6 Luna and GPT-5.6 Terra by default and not its best models.
A senior engineer shared a two-line config fix that routes memory through GPT-6 class models instead.
Step 1: Open Codex’s config file by typing this command:
open ~/.codex/config.tomlStep 2: Find the memories section (add it if missing) and pin the upgraded models.
[memories]
generate_memories = true
use_memories = true
extract_model = "gpt-6-luna"
consolidation_model = "gpt-6.1-sol"Step 3: Restart Codex. The “extract_model” setting now handles per-thread extraction, while “consolidation_model” manages the global merge. As a result, your new memories will become noticeably sharper over the next few sessions.
One trade-off is that bigger models use more of your rate limit during background generation. You can find the full list of settings in OpenAI's config reference.
IN CASE YOU MISSED IT
Our most-clicked story from yesterday
A former senior Meta engineer debunks the “software is dead” hype. Instead, he predicts a three-part future for software.
Whenever you're ready to dive deeper
If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:
We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.
Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.
What did you think of today's newsletter?
You can also reply directly to this email if you have suggestions, feedback, or questions.
Until next time — The Code team
