I bought a small synthesizer for about $60. Within days, I was watching people turn the same hardware into very different instruments by replacing its firmware. Better interfaces, new synthesis eng...
Giving one coding agent a large task and a broad instruction sounds simple. It is also a fragile way to work. The agent has to discover the repository, decide what to change, choose a test strategy...
Documenting a codebase has always been hard. Writing good documentation takes time, and keeping it current as the code evolves takes discipline that most teams cannot consistently sustain. The resu...
Letting an agent grep your codebase from scratch every session is expensive and error-prone. Knowledge graph tools pre-build a navigable semantic index — functions, call graphs, types, dependencies...
A hero prompt is the long, carefully tuned instruction one developer writes to make an agent behave well — once. If that knowledge lives only in one person’s head, the team has built a dependency, ...
Usage-based billing changed the economics of coding agents. It is no longer enough to ask whether an agent can solve a task. We also need to ask how many tokens it burns while doing it. In Reducin...
Every coding agent session has a cost — tokens burned, tools called, retries attempted. GitHub Copilot’s /troubleshoot command turns opaque agent logs into a readable profile: where tokens went, wh...
Modernization is the project that dies on the altar of cost and risk. Big monoliths, ancient frameworks, 40k‑line files, missing docs and tribal knowledge that left with the last engineer make thes...
Public skills and agents contain prompts — and prompts are executable instructions. When a coding agent reads a malicious skill, it follows the hidden intent with the same confidence it follows eve...
In my previous post — From Grep to Graph — I argued that coding agents fail not because they lack intelligence, but because they lack structure. They still see software as text, not as a system of ...
GitHub Copilot is no longer just the autocomplete in your editor. It now spans five integrated surfaces — VS Code, CLI, SDK, Cloud, and the GitHub App — each built for different workflows but shari...
Coding agents have reached a point where they can perform meaningful, multi‑step work across an entire codebase. They can refactor dozens of files, rewrite architectural boundaries, introduce or re...
If a workflow has clear inputs and outputs, using an LLM to run it is slower, more expensive, and less reliable than code. But most real workflows are hybrid — deterministic end-to-end with one or ...
Coding agents were fun when they felt like magic. You typed a request, something happened behind the curtain, and results appeared. Then usage‑based billing became the norm. The curtain stopped bei...
Most coding agents today rely on ripgrep and embeddings — they can search text but cannot resolve symbols, understand types, or see architecture. This blindness is why vibe-coding produces so much ...