The future of agentic engineering with Fletch.
Agents are changing who writes the code, not what it takes to build software that holds up. Writing on engineering in the age of agents, and how we're building Fletch for it.
Git worktrees are not an isolation boundary for coding agents
Most tools that run AI coding agents in parallel give each one a git worktree. A worktree shares refs, config, stash and hooks with your real repository. Paste-and-run repros showing an agent can execute code in your main repo and rewrite your commit identity, plus benchmarks showing a properly isolated clone costs the same.
Your AI agent doesn't need your GitHub token (or your .env)
AI agents rarely need the secret itself, only the action it authorizes. Least-privilege rules for GitHub tokens, API keys, and env vars, plus how Fletch keeps credentials out of the sandbox.
Fork the agent, not just the code: branching AI coding sessions
Exploration is where coding agents shine, but the usual ways to try a second approach throw away context. Branching an agent has two axes: code and context.
Running coding agents in Docker: when a write sandbox isn't enough
A write sandbox stops an agent from trashing your repo but still lets it read and reach the network. Here's what a container buys you, what it costs, and the credentials gotcha nobody warns you about.
Stop re-pasting your system prompt: skills and custom agents
Everyone keeps a graveyard of prompts they re-paste into every AI session. Split them into on-demand skills and reusable agent presets, and the drift stops.
How to actually review AI-generated code (you can't read it all)
At agent throughput you cannot line-read every diff. Here is a layered system for reviewing AI-generated code: automated review, tests as a gate, and human eyes where they matter.
Deterministic workflows for non-deterministic agents
LLMs give different output every run, and that is fine. The real problem is a non-deterministic process. Here is how to wrap a stochastic worker in a deterministic harness.
Claude Code vs. Codex vs. Cursor Agent vs. OpenCode: stop trying to pick one
An honest comparison of the four main AI coding agent CLIs, and why the smart move in 2026 is choosing a harness per task instead of marrying one.
Is it safe to run AI agents with permissions off? Sandboxing, explained
Almost everyone runs coding agents with permission prompts off. Here's the real threat model, why scale makes it worse, and what a write sandbox actually protects.
How to run multiple AI coding agents in parallel (without them stepping on each other)
A practical guide to running Claude Code and other AI agents in parallel: where terminal tabs, cp -r, and git worktrees break, and the principles that fix it.
Vibe coding vs. agentic engineering: generating code isn't shipping it
Vibe coding is fine for prototypes and dangerous in production. The fix isn't less AI, it's process: a plan, a review, tests, and a human at the merge gate.
Why I started building Fletch
The bottleneck in AI coding stopped being code generation. It became trust and review, and that is the problem Fletch was built to solve.