A recap of what caught our attention at Kaiku this week.
OpenAI’s DevDay announcements
OpenAI’s DevDay 2026 recap was packed with interesting developments. The Decisions API that looks a lot like the small decision models we have been experimenting with last week and this got the most attention from us. If you're also interested, all that is needed is the OpenAI subscription has an easy way to try the Decisions API.
We're also itching to try out Dots, always-on agents in the same family as Hermes and OpenClaw (which we have already been experiementing with). What we really are hoping for is much easier setup and secure deployments.
On the flipside of OpenAI related news was The Guardian reporting that OpenAI scrapped the planned release of Astra 6.1 after internal testing caught unwanted behaviour. Information is limited as always, but still the rate of change currently happening never cease to amaze.
One million agent skills
Absolutely worth a read is Vercel’s State of Agent Skills report, which says the skills.sh registry reached one million skills in seven months and nearly 280 million installs. A few numbers stood out:
- The top 375 skills account for 62% of all installs.
- Software engineering skills are oversupplied by about 30%, while business operations, writing and cloud infrastructure skills get far more installs per listing than average.
- Industry-agnostic skills make up 87.5% of installs.
The closing line sums it up well: the first million skills taught agents what everyone knows, and the next million will teach your agents what only you know.
The teams that compound context win
Tessl’s context flywheel argues that models are commoditising and tools are converging, so the lasting advantage is accumulated context. Running the cycle of generating, evaluating, distributing and observing context again and again makes agent output better and shows up subtler gaps each time.
The part that resonated most is ownership. Context needs maintenance, enablement and governance, or it decays like documentation. That mirrors closely what we have also been building towards ourselves with systems to manange the context flow, storage and retrival.
Agents move into Slack
Finally this week, Slack’s new code channels let you mention an agent to spin up a temporary channel, invite teammates and watch the work happen with approval before anything is deployed. After couple of weeks with Claude Tag, we found the output hard to review in chat, so this could help. For day-to-day work, though, a desktop app or the Claude Code CLI is still much easier to use.
Want to talk about where AI-native delivery is heading? Get in touch.
