The coding tools are eating each other. Claude Code is shipping background subagents. Codex burned through user quotas so fast it had to issue emergency refills. And the plain web version of Claude is apparently still better than all of it for writing a coherent sentence. We're watching the agent era arrive in real time, and the friction is very much on display.
01
OpenAI's Codex quietly burned through user quotas faster than anyone expected
Thibault Sottiaux, who works on Codex at OpenAI, posted a candid breakdown after users reported their AI coding budget disappearing far too quickly. The culprit was a cascade of small issues: auto-review had gotten more proactive than intended, a separate change was triggering more subagent work behind the scenes, and background suggestions were consuming credits users didn't realize were being spent. OpenAI reset usage limits and credited users an extra reset as a make-good. ---
Why it matters: This is what happens when agentic systems start doing more work autonomously. The thing you asked for runs, and then five other things run because the agent decided they were related. Your AI coding bill is now only loosely connected to the work you consciously assigned. Every team running Codex at scale should audit what's actually firing in the background, because the meter is running whether you're watching or not.
Claude Code is about to let subagents run while you're still talking
Boris Cherny, who works on Claude Code at Anthropic, announced that the next version will run subagents in the background by default. Right now, if Claude spins up a helper task, you wait. Soon, you keep working while the subagents handle whatever you sent them to do. ---
Why it matters: Pair this with yesterday's Codex story and you see the pattern. Both tools are moving toward agents that work in parallel without your attention. That's genuinely faster. It also means the Codex quota problem is almost certainly coming to Claude Code too. Background work is invisible work, and invisible work still costs money.
Microsoft Research wants to train your AI agent's instructions like a model
Microsoft Research's SkillOpt paper, published Monday, reframes how agent "skills" (the written instructions that tell an agent how to do a task) get improved. Right now, those instructions are mostly hand-tuned or generated once and left to drift. SkillOpt treats the skill file itself as something to optimize, running it through a training loop with validation checks, rejected-edit memory, and controlled edits, all without touching the underlying model. Across 52 evaluation tests spanning six benchmarks and seven models, SkillOpt was best or tied-best every time. ---
Why it matters: The agents your team is building probably have instructions someone wrote once, tweaked twice, and declared good enough. SkillOpt points at a real problem: those instructions silently degrade over time, especially as the agent gets updated around them. If this approach ships into production tooling, the question of "why did my agent stop working last week" gets a much better answer than "someone changed a prompt."
Coding agents make worse writers than the plain model, and someone finally said it
Peter Yang, who writes the Creator Science newsletter, noted that for writing and editing, basic Claude on the web still beats both Codex and Claude Code. His hypothesis: the system prompts that make coding agents good at code make them worse at prose. This is a small observation with a non-small implication. Every AI tool is optimized for its primary use case, and that optimization comes at a cost somewhere else. The instinct to route all AI work through your coding agent because it feels more powerful may be quietly degrading the quality of anything that isn't code. ---
Latent Space's AI news roundup: Arena hits $100M ARR, Cursor ships iOS, Meta decodes brain signals
The Latent Space newsletter's recap of June 27-29 surfaced a few items worth flagging. Chatbot Arena, the platform that lets humans rate AI models head-to-head, hit $100M in annualized revenue eight months after launching its paid evaluation product. Cursor shipped an iOS app with cloud agents that can run remotely on your desktop. And Meta announced Brain2Qwerty v2, a real-time system that decodes sentences from raw brain signals, with training code for both versions being released publicly.
Why it matters: Arena crossing $100M ARR on AI evaluation is the number that sticks. Someone has to be the independent referee for which models actually work, and the market is clearly willing to pay for that. When the labs control their own benchmarks, an independent scoreboard at commercial scale changes the incentives for who can fudge what.