Pere Villega

I want AI to give me time back.

Agent fluency can make an implausible number of projects feel possible.

An AI-native organisation is not a conventional team with more licences.

Formal methods have traditionally been expensive specialist work.

Once agents produce most implementation code, the durable asset is the factory around them: specifications, tools, feedback, state, evaluation, and recovery.

A human can supervise several agents, but attention does not scale with token spend.

A few shell scripts can coordinate a small fleet surprisingly well.

An overnight agent loop sounds like unattended productivity.

Agent teams can divide by task, role, or stage of work, but every extra communication path adds failure modes.

Parallel agents need more than separate terminal tabs.

Guest post by David Morgan-Brown on agentic reflective practice.

Adding a second agent is not a linear productivity upgrade.

A capable model is only one component of a dependable coding agent.

Line-by-line review does not scale with agent output, but abandoning review is not the answer.

Faster code generation only helps when the rest of the system can reject bad work quickly.

When an agent writes the implementation, the specification becomes our main engineering interface.

The difficult step in agent-assisted development is not enabling an autonomy flag.

Coding agents work well because repositories provide structure, history and checks.

Agents do not abolish the work between an idea and a dependable production result.

Agents can produce working code faster than we can absorb it.

Boris Cherny's workflow is less exotic than its throughput suggests: plan first, run agents in parallel, give them reliable verification, and turn repeated work into shared tooling.

Coding agents are good at iterating towards a result, provided we define that result independently.

A sandbox limits what an agent can damage while it works; it cannot tell you whether the resulting code is safe to ship.

An agent does not need your entire knowledge base in every prompt.

Skills are Standard Operating Procedures the agent loads only when needed — progressive disclosure applied to AI context.

Load 84 MCP tools and 15,540 tokens are gone before you ask a question; after thirty minutes you've burned 40% of your context on tool definitions you didn't use.

The context window is a budget, not a feature.

AI agents amplify whatever engineering process they're given.

A curated configuration for Claude Code that bundles plugins, skills, CLI tools, and sandbox guards into a single install script.

Your prompt is 0.1% of what the model sees; the other 99.9% is context engineering.

A complete guide to provisioning and hardening a Hetzner VPS for AI coding agents — covering cloud-init security, Tailscale mesh networking, Cloudflare Tunnel, Claude Code resource planning (including the subagent memory pitfalls), and automation scripts…

Skills 2.0 unifies custom commands and skills, adds subagents, evals, and a growing ecosystem.

Claude Code has shipped more features in four months than most tools ship in a year.

Every Claude Code session starts from scratch.

CLAUDE.md is the single file that determines whether your AI coding agent shines or flounders.

The setup checklist I wish I'd had on day one with Claude Code — isolation and sandboxing, terminal configuration, LSP integration, token monitoring, and what not to install.

A rigorous study found AI-assisted developers are 19% slower yet believe they're 24% faster.

AI coding agents have made code production nearly free.

The journey from AI-sceptic to agent-native isn't a smooth ramp — it's a series of uncomfortable jumps.

How I used Claude Code to build a GitHub backup tool I'd been putting off for years, and why LLMs are most valuable for the small internal tools you never get around to writing.

Why existing sandbox solutions for AI coding agents didn't work for my workflow, and how I built sandbox-claude: fully isolated Incus containers with btrfs snapshots, deploy key automation, egress filtering, and parallel agent support on macOS and Linux.

Why the 'junior developer' metaphor for LLMs creates dangerous expectations, and how the Monkey Paw framing from folklore leads to better decisions about autonomy, specification, and organisational governance.

Why the real value of AI tools isn't single-step task solving but tackling entire classes of problems at scale, from managing information overload to continuous monitoring, and how to find the leverage worth closing.

How LLM-generated slop PRs and shrinking small libraries are reshaping open source, eroding hiring signals, and making deep domain expertise the most defensible way for developers to stand out.

A TIL on using the --chrome flag with Claude Code CLI to enhance Claude in Chrome with custom skills, commands, and a tailored system prompt for more powerful browser automation.

Last week, thanks to David Morgan-Brown and Patchwork Health, I had the opportunity to do something I've genuinely missed: presenting in front of a live audience.

Durable computing has emerged as a topic in systems design.

There's a trend that has been going on for a while now.

I've written a new series of blog posts, a collaboration with Chaordic, that you can read in their blog: Understanding Data Mesh.

I've written a new blog post, a collaboration with Chaordic, that you can read in their blog: Running LLMs on your own.