Agents That Never Sleep: What August's Coding Tools Actually Changed

Spent a chunk of this month reading through vendor changelogs so you don’t have to, and the pattern that emerged wasn’t “new model, slightly better benchmark.” It was infrastructure. Everyone’s building for agents that run unattended for hours, not the polite little back-and-forth chat pattern we all started with. That’s the actual story this month, and it’s more interesting than any single release.

The one I keep coming back to is Claude Code flipping auto mode to the default. You now write permission rules as plain sentences instead of glob patterns, which sounds like a small UX nicety until you realise what’s underneath it: per-session caps on subagent spawns, MCP tool calls over two minutes getting shunted to the background automatically, /fork turning into its own background session. Anthropic is quietly building for the world where Claude Code is a thing running in a tmux pane on a server somewhere, not a thing you’re staring at. I run a version of this already for some overnight refactor jobs, and I’ll be honest, the permission model change makes me nervous in a good way. It means I actually have to think about what I’m authorising rather than clicking through defaults. Worth doing that audit before your next long session, not after something eats your repo.

MCP going stateless is the quieter but more consequential change. The protocol drops the initialize handshake and makes every request self-contained, so a plain round-robin load balancer can route it without sticky sessions. If you’ve ever had to fight session affinity on a hosted MCP server at 11pm, you’ll understand why this is a genuine improvement rather than protocol churn for its own sake. The catch, and there’s always a catch, is migration debt. Hardcoded -32002 error handling needs updating, and anything built against Roots, Sampling, or Logging is on a twelve-month deprecation clock. Budget a sprint. Nobody loves this kind of work but it’s the unglamorous stuff that actually matters once these tools are load-bearing.

Cursor’s cloud agents picking up event subscriptions and a persistent /goal command is the same story from a different vendor. Agents that wait for a Slack thread or a PR comment instead of being polled or re-invoked manually, that’s a real gap closing. I’m sceptical of how much of this is genuinely useful versus impressive in a demo, so I want to actually run /goal against a real repo before I say anything definitive. Filed under: will report back.

Then there’s Google, doing Google things. Antigravity getting folded into Gemini Enterprise licensing is being sold as governance and spend controls, which, fine, enterprises do want that. But The Register’s read is the honest one: Google is consolidating an agentic coding tool that started life inside DeepMind under its Cloud sales motion. If you value that tool’s independence, this is the direction of travel to watch, not celebrate.

Meanwhile Gemini 3.5 Pro has now missed three release windows since being announced at I/O in May. No model ID, no pricing, allowlist requests going unanswered since late July. I don’t know what’s happening internally at Google but three missed deadlines is a pattern, not bad luck. Compare that to Anthropic’s steady monthly Claude Code cadence and the contrast says something about discipline, even if I can’t tell you what it is exactly.

Small bright spot: Gemini 3 Flash landing in Gemini CLI at 78% on SWE-bench Verified, cheaper and faster than the Flash before it. For the boring high-frequency stuff, small refactors, lint fixes, PR triage, this is worth switching to and actually testing against your own repo rather than trusting the benchmark number outright. Flash-tier models have a history of trading multi-file reasoning for speed, and SWE-bench doesn’t always catch that.

On benchmarks generally: Claude Opus 5 leads Terminal-Bench 3.0 at 42.7%, but the gap to GPT-5.6 and GLM 5.3 has genuinely narrowed, and open-weight models like Qwen3.8 Max are now beating proprietary frontier models on SWE-bench Pro in specific categories. I don’t think there’s a clean “best model” story left to tell. Task fit and cost matter more than leaderboard position now, which is a more honest way to make decisions but a less satisfying one to write about.

None of this is revolutionary on its own. But add it up, stateless MCP, unattended Claude Code sessions, event-driven Cursor agents, and you get a picture of tools quietly being rebuilt for a world where the agent runs while you’re asleep, not while you’re watching. I find that genuinely useful and mildly unsettling in roughly equal measure, and I’m not going to pretend those two feelings resolve into anything tidier than that.