Here are the links and tools that caught my attention in March.
The orchestration explosion
I’ve been doing a lot of research around Agentic Engineering.
OpenAI’s Symphony, which turns project work into isolated, autonomous agent runs. It monitors task boards, spawns Codex agents, and lands PRs with proof of work. The reference implementation is in Elixir, which is a fun choice. A community fork called Better Symphony already exists that polls Linear, GitHub Issues, and GitHub PRs for work items.
OpenAI’s “Harness Engineering” describes a team that shipped roughly one million lines of code across 1,500 PRs in five months with zero manually written code. All Codex agents. 10x speed. The engineer’s role shifts entirely to designing environments and feedback loops.
Warp Oz does cloud-based orchestration for unlimited parallel agents and they claim 60%+ of Warp’s own merged PRs are created by Oz. Superset runs 10+ parallel CLI agents, each in an isolated git worktree. Conductor from Melty Labs runs parallel Claude Code and Codex agents on your Mac. OpenRouter Spawn puts AI agents on any cloud.
I also saved GitHub Spec Kit, a toolkit for Spec-Driven Development where specs become executable, and ProofShot, which gives AI agents eyes to verify the UI they build. Both feel like necessary plumbing for a world where you’re managing agents instead of writing code.
Speed vs. patience
“Good Software Knows When to Stop” is a satirical piece where ls gets replaced by “AI-Powered Directory Intelligence.” The point is real: good software resists scope creep.
Armin Ronacher (the Flask creator) wrote “Some Things Just Take Time” about how the speed obsession is eroding trust, relationships, and mature software. He calls the output of the rush “vibe slop.”
Garry Tan’s “Boil the Ocean” is the opposite take: retire “don’t boil the ocean” in the AI age.
Detecting AI prose
I saved multiple resources this month about recognizing and avoiding AI-generated writing.
“LLM Prose Tells” is a comprehensive catalog of LLM writing patterns — em-dash pivots, “delve,” hedge stacks, the whole catalog. It includes a copyediting checklist for de-slopping your text. The meta-funny part: it was written by Claude and then iteratively de-LLM’d.
Tropes.fyi provides a downloadable markdown file you can drop into system prompts to avoid AI writing patterns. Wikipedia’s Signs of AI Writing is a field guide for detecting AI content. I already have a detailed writing style guide that says “no AI slop” and “no em dashes.”
Breaking up with Software
Someone filed a complex 42-page return for free using Claude and IRS Free File. The companion Tax Assistant Skill is a Claude Code skill for tax prep. Intuit has spent $45M lobbying against free filing, and here’s a markdown file doing the job.
“Are You Just a .md File?” is a satirical SaaS survival scanner that assigns “death scores” to companies that could be replaced by a Claude skill.
Other finds
25 Years of Eggs: someone scanned 11,345 receipts dating back to 2001 and used AI to extract and analyze 25 years of egg purchases. Claude turned out to be the best OCR tool for the job. This is the kind of unhinged data project I respect.
I’ve moved from iTerm to cmux as my primary terminal, with a bit of Conductor as well. Became addicted to building with gstack.
Browserbase: Web browser for AI agents.
Google Workspace CLI: CLI for Drive, Gmail, Calendar, Sheets, and more.
Linear Method: Linear’s opinionated project management guide.
Remotion: Make videos programmatically.
Sprites: Stateful sandboxes.
Stripe Projects: Provision and manage services from the CLI.
Oboe: AI learning platform.
DataFast: Marketing channel attribution. Could be interesting for Hello Gravel.
Minifeed: Blog discovery.