In agentic systems, the principal, action, and resource are often unknown until the moment an agent acts, and the context that governs the decision arrives as a flood of signals from many sources. This post looks at how that context takes shape, where each signal is actually consumed, and why a non-directed world of agents still needs decisions that humans can inspect and predict.
Over the last three or four months I have been on a bit of a quest. I have been looking for a private digital assistant. Not another chatbot or the latest frontier model, but something I could run on my own machine. Something that would genuinely support me whilst keeping my data under my control. It has been a busy time with work and other commitments, but I managed to carve out an hour here and fifteen minutes there to explore some of the more exciting developments in AI. I think I have finally found what I was looking for. It is not some amazing secret project. Far from it. However, getting to this point involved a surprising amount of experimentation, frustration, and the occasional dead end. In this series I intend to share the challenges, discoveries, successes, and inevitable setbacks that I have encountered, in the hope that my experiences and insights might help some of you avoid the same rabbit holes.
Agents that narrate their work are the best team players
Agents like Claude Codex and Codex run in your terminal. From that strategic vantage point they can wield system tools (awk, bash, curl, git, python) as well as MCP tools. Here’s how that loo…
Obsidian Vault to Enterprise Company Brain: Where Agents Write, Branch and Merge
From a personal Obsidian vault to an enterprise company brain: how agents write context on versioned branches, typed graphs keep it consistent, and merge.
Can one agent start another? Queueing sub-tasks in GitHub Agentic Workflows
Consider a "dark factory" setup: a software factory with almost no people on the production line. A small number of people orchestrate the system, but agents do most of the work: planning, coding, reviewing, securing, deploying and reporting back. The goal is to define the factory itself as code.
How To Build Your Own LLM Runtime From Scratch | Towards Data Science
Demystifying the LLM runtime: A from-scratch tutorial on building a custom C++/CUDA inference engine, understanding how bare-metal AI actually works, and meeting the synchronization bugs on the way.
Vector RAG Isn’t Enough — I Built a Context Graph Layer for Multi-Agent Memory | Towards Data Science
A structured context graph beat a flat chat dump and a vector-only RAG pipeline on the same multi-agent conversations — full working code, real benchmark numbers, zero API calls
Generative AI in the Real World: Local Voice AI with Pete Warden
Pete Warden has spent his career on the frontier of small, local AI, first as one of deep learning’s earliest engineers (he coined the term “TinyML”) and now as founder of Useful Sensors and Moonshine AI, where he builds voice models that run entirely on-device. Pete joined Ben to make the case that local AI …
What a User Story Actually Costs in a Dark Code Factory
Between March and July 2026, I built a production application of 861,601 lines of code. This included 696 user stories and 779 merged pull requests over 105
NOPE measures how conversations between people and AIs are going, for both parties, and publishes what it finds: benchmarks, incident records, open-weight models.
