The AI coding tools story this week keeps circling the same uncomfortable truth: these things are consuming resources faster than anyone planned, and the people building them are figuring out the limits in real time. Boris Cherny's thread from yesterday about job titles blurring is starting to look like foreshadowing. When your coding agent is silently spinning up subagents in the background while burning through your monthly quota, "who's responsible for this" is no longer a philosophical question.
01
Claude Code is about to run several jobs at once while you're still asking it questions
Boris Cherny, who works on Claude Code at Anthropic, announced that the next version will run subagents in the background by default. Instead of waiting for one task to finish before starting the next, Claude can keep a conversation going with you in the foreground while other agents work in parallel behind the scenes. You can still override this and run things sequentially if you want. ---
Why it matters: This is what "agentic" actually looks like in practice. Not one AI doing one thing, but a cluster of processes running simultaneously, all on your quota, all potentially interacting in ways you can't see. That background activity is exactly what tripped up OpenAI's Codex users this week.
OpenAI apologizes after Codex quietly ate users' monthly allowances
Thibault Sottiaux from OpenAI posted a postmortem after users reported their Codex usage limits draining faster than expected. The culprit was several compounding issues: auto-review had become more proactive than intended, another change was triggering extra subagent work, and background suggestions were running more than they should. OpenAI is resetting limits and crediting an additional reset to affected accounts. ---
Why it matters: The pattern here matters more than the specific bug. As coding agents do more automatically, the line between "work I asked for" and "work the agent decided to do" gets blurry. If you're managing a team budget around AI usage, you now have a new line item to watch: background agent activity you didn't explicitly request.
Microsoft Research built a system to stop AI agents from getting worse every time you edit their instructions
The core problem with AI agents in production: every time a human edits an agent's instructions to fix one thing, there's no guarantee something else doesn't break. Microsoft Research's SkillOpt treats an agent's instruction file like a trainable parameter, running the editing process like an optimization loop rather than a one-off rewrite. It tests changes against held-out validation, remembers which edits failed, and keeps instructions compact enough for humans to audit. Across 52 evaluation setups covering six benchmarks and seven models, it was the top or tied-top approach in every single one. ---
Why it matters: Every company that has deployed an AI agent has a person who "maintains the prompts." That person currently has no systematic way to know if their latest fix made things better or just shifted the failure mode. SkillOpt is the first credible answer to that problem at research scale, and Microsoft will push it toward Azure customers fast.
Peter Yang: coding agents are bad writers, and the reason is probably their system prompts
Peter Yang, who writes about product and AI, posted a short observation that landed: plain Claude on the web is still better for writing and editing than Codex or Claude Code. His hypothesis is that the coding-specific system prompts tuned into those agents make them worse at prose. ---
Why it matters: If you're using your coding assistant to draft documentation, emails, or product specs because it's already open, you may be getting worse output than you'd get from the same underlying model without the coding context wrapped around it. Tool choice is not neutral.
The Latent Space newsletter on what happened in AI while everyone watched Germany lose
The June 27-29 roundup from Latent Space caught a few items worth flagging: Meta released Brain2Qwerty v2, a real-time decoder that turns raw brain signals into text, along with training code for both versions. Cursor shipped an iOS app with always-on cloud agents that can control agents running on your desktop. Arena, the AI evaluation platform, hit $100M ARR eight months after launching its commercial product. And Cline launched a $9.99/month pass for discounted access to a bundle of models including DeepSeek, Qwen, and Kimi.
Why it matters: Arena hitting $100M ARR on evaluation tooling tells you something real about where enterprise AI spending is going. Companies are no longer just buying models. They're buying the infrastructure to figure out whether the models they already bought are working.