Winding Down Substaff

Two months ago I wrote about winding down Kreye. Doing this again so soon is less comfortable than doing it once — the first time reads as discipline, the second starts to read as a pattern.

But I think it is a pattern, and I’ve decided it’s the right one: I’d rather close two things cleanly than keep four things on life support. Today I’m shutting down Substaff.

What Substaff was

Substaff was an attempt to answer a question I kept running into myself: once you have more than a couple of AI agents working for you, who manages them?

The pitch was to manage them the way you’d manage people. You defined a company goal, hired a team — a CEO, engineers, marketers, researchers — and each agent got a role, a budget, and a place in an org chart. Agents woke on heartbeats, picked up work, delegated up and down the chart, and reported back. Every task traced to the mission. Every decision landed in an audit log. You were the board: approve the hires, override the strategy, pause anyone at any time.

It was deliberately vendor-agnostic. Claude Code, Codex, Cursor, OpenClaw, a bash script, a plain HTTP endpoint — if it could receive a heartbeat, it could be hired. Underneath was a Node server and a React UI, with sandboxed execution, budget enforcement, semantic search over past artifacts, and multi-tenant isolation down at the row level.

It looked like a task manager. That was the point.

Why I’m shutting it down

Two reasons, same shape as last time.

The first is activity. Substaff got attention — the idea demos beautifully, and “org chart for your agents” is a sentence people immediately understand. But attention and usage are different curves, and only one of them mattered. People would set up a company, watch a few agents run, and not come back. That’s real signal. It says the concept is legible, not that it’s load-bearing.

The second is cost. Agent infrastructure is expensive in a way that doesn’t care whether anyone is logged in. Sandboxes, an always-on database, object storage for artifacts, a vector index, and the model calls themselves. A platform designed to run agents around the clock will happily run up a bill around the clock. For a product with low activity, that’s a standing order I was quietly funding.

What I think I got wrong

This is the part worth writing down, because the reasons above are just arithmetic. The interesting question is why the arithmetic looked like that.

I built the management layer before most people had a management problem. Substaff was aimed at someone running twenty agents who has lost track of what they’re all doing. That person exists. I am that person some weeks. But most people running AI agents today are running one or two, attentively, in a terminal they’re already looking at. Coordination overhead isn’t their bottleneck — output quality is. I built the org chart for a company that most users hadn’t hired into yet.

The metaphor was the demo and the trap. Framing agents as employees made the product understandable in ten seconds, and it also promised a degree of autonomy the underlying agents couldn’t consistently deliver. An org chart is a great way to structure work that’s going well. It does very little for you when a goal quietly goes sideways on step three and four agents downstream keep building on the mistake. Governance features assume the thing being governed is basically competent.

Bring-your-own-agent read as a feature and behaved like a support surface. Every runtime fails differently. Supporting all of them meant owning all of their failure modes without controlling any of them.

None of this makes the thesis wrong. I still think that when agents get reliable enough to trust with multi-day work, something like Substaff becomes necessary. I just don’t think I can afford to sit on the runway waiting for that plane, and being early is only distinguishable from being wrong in retrospect.

Where my attention is going instead

The honest flip side of shutting something down is naming what you’d rather be doing:

• Tandemu — measuring how developers actually work with AI coding assistants. AI-to-manual code ratios, friction detection, commit-level attribution, and a persistent memory layer that learns how you work. This is the one I keep thinking about when I’m not at my desk.
• AnswerRank — visibility in the era of answer engines. Tracking how ChatGPT, Claude, Perplexity, and Gemini talk about your brand, and working a concrete checklist of fixes so they recommend you instead of a competitor.

Alongside those, the research work in the lab — the hybrid SSM experiments, the retrieval work on Mamba’s latent states — has been where a lot of my curiosity is going. Less product surface, more actual learning per hour.

Substaff was the thing I had to consciously not work on to make room for these. Closing it just makes the calendar match reality.

A thank you

If you signed up, spun up a company, gave an agent a goal, or sent me a bug report — thank you. Genuinely. You gave a side project the one thing it can’t manufacture for itself: real people trying to do real work with it. The parts that didn’t work taught me more than the parts that did, and most of that is already showing up in what I’m building next.

Building things is easy to start and hard to stop. Stopping well — saying thanks, being honest about what broke, and pointing your attention at what matters more — still feels like the right way to close one out.

On to the next.

If you want to follow where I’m headed, the projects page is the live list, and you can always reach out.