Here are the links and tools that caught my attention in May.
At Hello Gravel I’ve been working on how agents fit into our day-to-day, including making some of our internal systems reachable by API so Claude and ChatGPT can do real work in them. Most of what I saved in May relates to the same question: what does it look like when a real company, not a demo, runs on agents?
AI-native is a spectrum, not a switch
Floodgate co-founder Ann Miura-Ko has been visiting companies, most recently Ramp, and argues that “AI-pilled” gets used like it is binary when it is not. She borrows the self-driving levels: for years everyone chased Level 5, and the levels mattered because they forced precision. Cruise control is not autonomy. Most companies calling themselves AI-native are running cruise control. At a five-person startup, that is fine because the founders are the operating system. At scale, AI has to become part of how the company works rather than a personality trait of the founding team.
Ramp has written useful field reports on closing that gap. They explained why they built their own background agent, and Seb Goddijn described building every employee their own AI coworker, a suite they call Glass. One detail: they hit 99% adoption and then noticed most people were stuck, using the models “like driving a Ferrari with the handbrake on.” It was not a lack of ambition. Terminal windows, npm installs, and MCP configs were a wall. The blocker had moved from the model to the environment around it, and everyone who pushed through ended up with a different setup and no way to share what they learned.
Shopify CEO Tobi Lütke published a writeup on River, Shopify’s AI agent that lives in Slack. The design constraint is simple: River only operates in the open. No DMs. Ask in a private message and she declines, then asks you to move to a public channel. In the last 30 days, nearly 6,000 employees worked with her across 4,450 channels, and about one in eight PRs merged into Shopify’s monorepo was authored by River and reviewed by a human. The part I found more interesting than the numbers: people learn by watching each other work with her. Tobi connects it to dropping out of school at 16 and learning to program by making coffee for the good programmers in the basement until their judgment rubbed off. Most agent rollouts focus on throughput and skip the apprenticeship.
I saved a couple of platforms chasing this shape: Hyperagent (“the system of agents that does real work, learns how your organization operates”) and Mercury Intelligence, whose pitch is making any agent “become a team player.” The interface keeps drifting from chat box to org chart.
The harness is still the product
The other half of my reading was about directing these things. The .txt team wrote The bottleneck was never the code. They had an experiment that sat on the roadmap for over a year. It kept coming up in conversation and kept getting deferred. Then the author spent half an hour explaining the method to Codex, OpenAI’s coding agent, and had a working first version a few hours later. The blocker was everything around the code: the framing, the decision to start, the willingness to run the thing. I’ve run into the same thing more than once.
A lot of the tooling I saved is about formalizing that direction. Flue bills itself as an agent harness framework rather than another SDK, with the formula “Agent = Model + Harness.” OpenAI’s cookbook entry on using Goals in Codex adds persistent objectives that keep a thread working toward a defined outcome across turns. Garry Tan wrote up Meta-Meta-Prompting, his explanation for why the CEO of Y Combinator is coding until 2am. And the Claude Agent SDK workshop from Thariq Shihipar at Anthropic is on my list to sit through.
For running many agents at once, KanBots is a kanban board where every card dispatches one of eleven agent CLIs into its own git worktree, or you hit autopilot and let personas split the work while you sleep. I’ve mostly stuck to one agent, one project, one session, but the tools keep pulling toward fleets.
Two skill packs were worth a look. Dangerous Professional implements Patrick McKenzie’s (patio11) framework for writing firm, rights-aware letters to banks, insurers, and contractors. It reads your documents, finds your leverage, and drafts correspondence grounded in the institution’s own records. I have an insurance situation this is basically built for. Corey Haines published marketingskills, a pack of CRO, copywriting, SEO, and growth skills for agents. I spend enough time on paid-media math that I’ll try wiring some of these in at work.
HTML is the new markdown
Thariq had a second post arguing that HTML is the new markdown. He has stopped writing markdown files for almost everything and has Claude Code generate HTML instead, because the moment you want layout, color, or interactivity, markdown taps out and HTML does not. I generate so much markdown that I built a tool just to turn it into PDFs, so I get the argument.
Follow that thread far enough and you end up at Google’s HTML-in-Canvas work, which renders live HTML inside a canvas element. The demos are fun: a form that is also a Duck Hunt shooting game, buttons that ripple like cloth. No practical use for it yet.
The web-standards saves fit the same mood. The Website Specification is a platform-agnostic, MIT-licensed spec for what a good website should do, “from <title> to /.well-known/security.txt,” written for humans and agents. Anna’s Archive published a post literally addressed to machines: If you’re an LLM, please read this. We spent the last year teaching agents to read our websites, and now we are writing pages for them. This comes up at work too: we are in the middle of a website migration, and one open question is whether to remove bot blocking so the AI crawlers can read us.
One more in this neighborhood: Printing Press. Point it at an API spec, a website with no public API, or even a community fan project, and one prompt prints a token-efficient Go CLI, a Claude Code skill, an OpenClaw skill, and an MCP server. It bakes in Peter Steinberger’s playbook: a local SQLite mirror beats a remote API call, and compound commands beat ten round trips. They call it “muscle memory for agents,” and it is close to a layer I’ve been building by hand at work.
The witness effect
Not everything I saved was about tools. Garry Tan wrote YC as a Witness, built on psychologist Alice Miller’s research into why some kids who grow up in brutal circumstances become whole adults and others repeat the cycle. Her answer: whether the child had at least one person who saw what was happening and believed them. Not someone who could fix it, just someone who could say: I see you, this is real, you’re not crazy. Garry argues that is most of what YC does. A founder walks in with an idea everyone in their life thinks is insane, and the room’s job is to be the witness who has seen what good looks like at the pre-everything stage.
On the practical founder-education side, Marc Lou’s CodeFast is a coding course aimed at entrepreneurs instead of software engineers. Their pitch: traditional courses teach you to “master 47 sorting algorithms you’ll never implement” when what a founder needs is a paying product live in two weeks. A useful reference for teaching people to build the first version themselves.
And Claire Vo wrote Earn your content, a counterweight to all of the above. Her target is a world where describing has the same perceived value as doing, where listening feels like learning and watching feels like building. I save a lot of links. The point of saving them is to go do the thing, not to feel like I did it by reading about it.
Stuff I bookmarked
A grab bag of lighter saves, heavy on the homelab this month:
- paperless-ngx: Self-hosted document management that scans, indexes, and archives everything. I have a paper problem.
- AdGuard Home: Network-wide ad and tracker blocking at the DNS level.
- Voice Terminal: Open-source, voice-first AI terminal for ESP32-class devices. Self-host the server, flash the firmware.
- OpenHome DevKit: An AI voice agent SDK and hardware kit. I have no project for it, which has never stopped me.
- Crowd Supply: Where engineers launch and sell hardware. A dangerous place to browse.
- PaceBar: A Mac menu-bar instrument that notices when your work pace is rising. Runs entirely local.
- HDIT: High Density Interval Thinking, a simple structure for working through one hard question with full attention.
- Exa: AI search API. They just raised a $250M Series C. On my list for agent web search.
- Stainless: Generates SDKs, MCP servers, and API docs from a spec in minutes. Relevant to the API work above.
- Strapi, Forgejo, and Grafana: A headless CMS, a lightweight self-hosted git forge, and observability. The own-your-stack cluster.
- The car dealer’s worst nightmare: WSJ on a guy who earns $1,000 a job negotiating car purchases over YouTube. A great one-person, distribution-first business.