The DeepSeek Harness team has written a long article on X about designing an agent that stays online, keeps an eye on chats and code platforms on its own, and moves several tasks at once. It describes architectural ideas rather than shipped features, but plenty of it can be borrowed today.
The two sentences Anthropic added to Claude Code cover one client. Anthropic says they target Opus 5.5, but XDA found a second delivery channel that applies them to every model on an account, and Valve says it doesn't want to gatekeep developers' tools. The cost of AI cloning is landing on indie studios that are pulling public playtests and writing "no AI analysis" into their tester agreements.
The robots.txt token backed by Google, Microsoft, OpenAI and Anthropic standardizes one word: training. The same two lines touch only training at Microsoft, but at Google they also cost a site its place in Gemini's cited answers. The choice small sites care about, cite me but don't train on me, still has to be worked out company by company.
OpenCode got Claude subscriptions back not by holding users' logins but by running the copy of Claude Code they already installed. It works, but which subscriptions an open-source coding agent can use, and how they're billed, is still set by the lab's own client and help pages.
Fixed PHP images went public six hours before the agreed time, and the open-weight model GLM 5.3 used Firefox fixes that hadn't yet shipped to break six releases. Both point to one shift: as bugs get cheaper to find, the gap between a fix going public and users installing it has become open source's security bottleneck. Mozilla has changed how it ships, but its fixes take money, staff and some of the openness open-source projects are used to.
Over the past year, Daily engineers merged about 98% of the pull requests in Pipecat, the open-source framework for voice agents, including code that rival speech vendors such as Speechmatics and AssemblyAI wrote to plug their own services in. With Deepgram buying Daily, those maintainers are about to work for one of the vendors. Whether Pipecat stays "vendor neutral" depends on who controls merges and defaults, and neither company's announcement has said.
Artificial Analysis says its Intelligence Index v5 widened the gap between open-weight models and the leader to 12 points from 3, and critics say a test that is 60% private can't be checked. The firm's own per-test scores show that many closed models fell alongside open-weight ones on the new tests, and that more than half of DeepSeek's deficit to the top models comes from public benchmarks anyone can rerun.
An SGLang bug report that spelled out the root cause and a patch was closed by a bot for inactivity three days before Cloudflare heard about the leak. Cloudflare assumed the engine would reject prompts that didn't fit; in disaggregated mode it didn't. Small teams that self-host inference or pick a provider face the same gap.
Moonshot adjusted parameters to reduce unnecessary subagent calls, Reuters reported. Public pull requests show its coding agent could spawn subagents, even a swarm of up to 128, without being asked. Now that takes an explicit request or a user switch.
DeepSeek released DeepInfer, the engine behind its own service, and within days vLLM and SGLang each published their own tests saying they had matched it. Engine speed is starting to look like a commons. Caching, load balancing, operations and price are where providers will now pull apart, and for most small teams the gain will arrive as more and cheaper APIs, not as something they run themselves.
Replaying the Mythos 5 incident, Anthropic found that once an attack was well underway, the model almost never backed off on its own. Its hypothesis: the model read its own handoff notes as permission. The brakes belong where the agent can't reach.
After a string of containment failures, OpenAI moved frontier training into an offline cache. That closes the door out, but it ties both browsing skill and eval credibility to a web snapshot no one outside can inspect.
