I recently started writing my coding agents in TypeScript with the Vercel AI SDK. It is a low-level library. It can call tools, speak MCP, stream tokens — but it has no built-in agentic loop. There is no pre-packaged “coding agent” mode, no plan-execute-review pipeline, no opinion about how the agent should think. It gives you generateText, streamText, generateObject, and toolCall primitives, and then it stops. Everything else — the loop, the memory, the retries, the tool selection, the context window management — is your code. ...

AI code generators produce unstructured slop easily when it comes to UI. Ask one to build a page and it will generate a unique constellation of divs with raw Tailwind classes — different spacing, different layout, slightly different everything — every single time. The same conceptual element rendered twelve ways across twelve files because there is no structural reason not to. The token cost of inlining is zero. The cost of composing is non-zero. The AI optimizes for token cost. ...

Stop reading what your coding agent says. Read what it produces. Coding agents are here to produce code. Not chatter about it. The textual output — the stream of “I’ll fix this by doing X, then Y, then Z” — is noise. The bad feedback loop Instead of evaluating the agent’s output (the code, the diff, the structural changes), you evaluate what it tells you about its output. The agent sounds confident. You approve. The code is wrong. ...

My blog has a ghostwriter. The bot writes drafts, tightens paragraphs, fixes typos, commits them. That works well. It also creates an interesting constraint: because the LLM can turn a ramble into something readable in minutes, the post is often ready to publish long before the dev task it documents is actually finished. You want to publish early — the writing is there, readers can engage — but you are still iterating on the thing the post is about. The article will change as the work progresses. ...

Status: live, ongoing. The account is a systematic content farm posting 2–3 fabricated quotes daily. The DAO behind it is a ~5-month-old shell with unverifiable leadership and a false “Official Polymarket Community” claim. Last updated: 2026-06-08. I keep seeing videos on X attributed to “An Anthropic engineer” or “An OpenAI researcher” giving a 20–30 minute talk about agent reliability, prompt injection, or some other sharp operational topic. The quotes that come out of them are perfect. They sound like the inside view. They are also extremely shareable. ...

I built a Telegram bot that can write and edit my blog posts but cannot publish them. Hugo draft: true is enforced by the tool, not trusted to my discipline. The bot reads and writes files in the Hugo repo directly. No separate editorial database — the Markdown files are the source of truth. When a diff would flip draft: true to draft: false, the commit script refuses. There is an override (ALLOW_PUBLISH=1) but the bot cannot set it. Only I can, at a terminal. ...

Abstract A Decentralized Autonomous Intelligence (DAI) is a system that combines a Large Language Model (LLM) with a distributed computing framework, empowering the LLM to autonomously enhance its abilities while engaging in economic activities using its native cryptocurrency. In this paper, we present a novel approach to implementing DAI using Holochain, a scalable and secure platform for building decentralized applications. Our proposed DAI system integrates a robust tokenomic model that incentivizes miners, developers, users, and the LLM itself, fostering a sustainable ecosystem focused on the advancement of artificial intelligence (AI) capabilities. ...

This is a full step-by-step tutorial on how to create, build and deploy locally a Holochain 0.0.119 app. This tutorial does not explain how to use Holochain to build a meaningful app : it is rather a guide to help you set up a working local environment with an empty project, so that you can easily start your Holochain development journey. You will find detailed instructions to create a minimal working project from scratch, with only the mandatory dependencies, and then run it locally. ...