Agents

When I first started building my personal AI assistant, Clawdia, she would always get the time wrong. It took a few attempts to fix. I tried prompting. I tried putting the timezone in her config. Eventually she built a hook that injects the current date and time into every single message before she even sees it. And then I forgot it was ever a problem. That’s kind of how working with an AI assistant actually goes. You hit something broken, you get the agent to fix it, it breaks differently, you prompt it to try and fix it again, and then at some point it just starts working and you can’t even remember when it crossed the line. Nobody writes about that part. They write about the setup, the first magical moment you see it just figures something out on its own, maybe the architecture diagram. They don’t write about the three weeks where your agent keeps suggesting cauliflower rice even though you’ve told it four times you hate it. ...

The biggest GPU company on the planet just built an enterprise wrapper around a free open-source tool because their customers are already using it — and IT can’t stop them. At GTC this week, NVIDIA announced NemoClaw — a software stack that deploys secure AI agents on OpenClaw with a single command. Enterprise guardrails, audit trails, sandboxed execution. It’s a good product, and it’s a confession. I talk to organisations every day that are trying to build enterprise agents. To be clear: the first wave of enterprise AI worked. Chatbots, RAG pipelines, hybrid search — organisations have genuinely nailed that stuff, and it’s delivering real value. But agentic is a different problem entirely. Outside of coding — which works, and works well — most organisations are struggling to build agents that actually do things autonomously on behalf of users in a way that changes how the business operates. The demos are fine. The production value is… not. OpenAI just declared a “code red” and pivoted its entire company toward coding and enterprise because that’s the one agentic use case that’s actually delivering. That should tell you something about where we are. ...

Five weeks ago I set up an AI agent called Clawdia using OpenClaw. I expected to play with it for a weekend and move on. She’s still running. She’s become the most useful tool I’ve ever used, and I don’t think I could go back. What OpenClaw actually is The TLDR is, OpenClaw is an open-source, self-hosted AI agent. You run it on your own machine, connect it to Telegram or WhatsApp, and talk to it like you’d message anyone. No dashboard. No app. Just a chat window. ...

When building AI agents, one of the biggest challenges is tool selection—especially as the number of available tools increases. Evidence from OpenAI’s developer community and technical documentation highlights that agents’ tool selection accuracy tends to drop noticeably with larger toolsets, sometimes causing irrelevant or random tools to be invoked even with carefully designed prompts. OpenAI’s internal discussions note a clear decline in single-tool call accuracy and an uptick in failure modes as the system prompt and available tool list grows, with developer complaints of incorrect or unexpected tool choices despite best practices. ...