“don’t train your own model” is common ai advice. it's wrong. your token bill's the proof. today, we’re excited to launch castform into open preview. castform is the easiest way for you to train your own model, on your own data. open-weights models are performant and much cheaper. when trained on your task & proprietary data, they beat closed models. the thing standing between you and that was weeks of plumbing & years of ml expertise. with castform, model training is as simple as prompt engineering. @castformai bring your agent traces or raw corpora. castform turns it into training data, picks the right algorithmic recipes, manages gpus, and gives you an ide to watch and chat with your model as it learns. see what you can build with castform👇

Unfortunately, this is inevitable. Any time an agent has a) outside exposure and b) tool use, bad actors will test its limits. We see attacks like this all the time at Runlayer, and we can stop them. It's not as easy as you'd think. x.com/zeeg/status/2062808524…

Don’t ask for permission

We've built 150 internal agents at Runlayer, and this one might be my favorite. Steal it. The Dropped Balls Catcher. It tackles one of the biggest pain points our team members face. They can't keep track of every thing they said they'd do. Slides they promised. Emails they said they'd follow-up on. PRs they said they'd review, but forgot about. One agent fixes this. It pulls from call recordings, emails, calendar invites, Slack, and Linear tasks to remind you of everything you said you'd do, but haven't done yet. Every day. Most people have this problem, and none of our employees have it anymore. If it's helpful, I'll drop the full prompt in a post later this week. I'm looking to give as much value away as possible here.

"Agent control plane" is starting to look like a real category. A few months ago the question was whether agents even needed a control layer. Now it's which one. Over the last few weeks a handful of products showed up around the same problem, seeing what agents are doing, scoping what they can touch, keeping a human in the loop when it matters. Not for some hypothetical future. Because agents are already wired into real systems and real data, today. When a bunch of companies start building the same layer at the same time, that's usually what a category forming looks like.

You can't make an AI-native company by mandate. When you try, 5% of people run with AI. The other 95% hit the valley of despair, and give up. Buying Claude licenses is the first step, but it's a smoke screen. Your employees hit the valley of despair next. - They want to connect Claude to sensitive tools and systems. - They want to build and share skill files. - They want to build agents. That’s the part that gets hard. Adding Claude doesn't magically change your workflows. Companies succeed by making the right way to use AI the easy way. Concretely: 1. Right path: Golden path for every employee, not just power users 2. Right platform: enablement AND control in one platform. Security platforms stifle adoption. They're not built for actual usage. 3. Right choice: a platform used by actual AI-native companies, not laggards. This is how we designed @Runlayer. This is how AI-native customers like Gusto and Instacart use Runlayer. Those executives mandated AI usage, sure. They also gave every single employee a golden path.

No one talks about this, but the world of agents is shockingly diverse

Tons of vibe metrics floating around for agent usage right now. - "This agent comes up a lot in meetings" - "Our best AI-native engineer says this one is good" - "I think our CEO told everyone to use this one because he made it" These are all awful ways to measure agents. We specifically built a feature in Runlayer to segment agents by cost. We tell you what's being used, what's costing too much per run, and what's underutilized. Get the hard numbers. Wasted token spend is almost as bad as no token spend at all.

Every new employee gets an AI agent when they onboard @Runlayer. You log in to your new laptop, and our agent creates a custom onboarding for you with 16 sections. 1. Role context 2. Interview recap 3. Platform overview 4. Product strategy & roadmap 5. How we work 6. Tech stack 7. Security model 8. Current customers & what's hot 9. Current engineering sprint 10. 30-60-90 milestones 11. First week playbook 12. Team directory 13. Slack channel guide 14. Must-read links from #share 15. Key resources 16. Connect Runlayer MCP to your AI client If you have questions, the agent surfaces any context you need from tickets or Slack threads, even from before you joined. It also sends a daily digest every morning at 9am with a bunch of information. All maintained in a custom Notion page. If your onboarding doesn't look like this, use this as a guide. Better yet, DM me and we can set this up easily with Runlayer.

It takes 6 words to build an AI agent with our Slack assistant. When your HR rep tags Runlayer in Slack and says "Build me an agent that pulls candidate notes from our ATS, summarizes weekly, and posts to #hiring on Fridays" there's a few things that need to already have happened. 1. We already linked to your ATS as an internal MCP 2. We already approved and vetted the skill the agent uses 3. We already linked the skill's permissions to the user's identity With us, it just works. Think about how hard that workflow is normally, though. That same HR person would have to (a) get IT to approve the ATS plugin, (b) hand-configure credentials on their laptop, (c) hope it works in Cursor or Claude, (d) re-configure it next time they switch machines, (e) be quietly violating security policy the whole time. This is why AI adoption stalls. I would rather quit my job than have to adopt AI through that workflow. The golden path to AI adoption is making agent creation a matter of words. That is how you create AI-native employees.

In the last two weeks: ServiceNow shipped Action Fabric, AWS MCP Server hit GA, Microsoft moved Agent 365 to GA. The agent execution layer is the new cloud. Ignore it now, pay for it in 2027.

sir, this is Azure console

MCP is necessary for the enterprise. CLI doesn't work or scale for management. APIs weren't built for agents. Model improvement doesn't fix prompt injections or 100s of other security issues. MCP secure control plane. I can't see it any other way. x.com/signulll/status/204968…

OpenAI rewarded creature metaphors while training one personality. The behavior leaked across every personality. Their fix: a system prompt that says 'never talk about goblins.' RL rewards don't stay where you put them. Neither do agent permissions openai.com/index/where-the-g…