I built the AI pivot from the inside. It didn't work.

In October 2024 a TypeScript framework called Eliza launched, and within weeks it was one of the fastest-growing repositories on GitHub, the most adopted agent framework in crypto almost overnight.¹ Suddenly every crypto company had an agent strategy. I was inside one of them: a Solana content-monetization protocol where I’d been since the founding week, nearly four years by then. We pivoted to AI agents, and I built a large part of that pivot with my own hands.

It didn’t work. The agents mostly died. And I would do it again, because that failure is the direct ancestor of everything I now build. Most pivot post-mortems are written by people who watched one from a conference seat, or cleaned up by people who need it to have been a strategy. This one is from inside, and I’d rather keep the embarrassing parts.

The wave, without the retcon

Here is the sentence that every pivot retrospective edits out, so let me write it plainly: we pivoted because it was the thing to do. The wave was real and enormous. AI-agent tokens went from almost nothing in mid-2024 to a category worth tens of billions of dollars by early January 2025², and the idea followed the wave, not the other way around. Agents for creators, connected to paid third-party data APIs. That was the pitch. Nobody was pretending we’d found it by interviewing suffering users.

I want to be precise about the honesty here, because it cuts two ways. The company’s motivation was the hype. Mine was adjacent but different: what pulled me in was the substrate. Agents that act need memory, tools, and somewhere safe to run: sandboxes, isolation, little VMs you can hand to a process you don’t fully trust. I wanted to build that. Mostly the exercise, if I’m honest. It turned out to be the only part of the pivot that compounded.

What we actually built

We started, like half the industry, from Eliza. And the first lesson, free of charge: the fastest-growing repository on GitHub was also a mess. There was so much bad code that I forked it and rebuilt the agent from the ground up in TypeScript, keeping the ideas and replacing the body. Adoption speed and code quality turned out to be completely uncorrelated, a thing worth remembering every time a framework’s star chart goes vertical.

Around the agent, with the team, a RAG and memory backend in Python. On top of it, the platform piece: users could spawn their own agent, wired into our backend (its memory, its retrieval, about ten provider integrations to the data APIs the pitch was built on).

Then, in March 2025, the piece I built alone and the one decision from the whole pivot I’d frame: an MCP server. Anthropic had published the Model Context Protocol that November,³ and I found it the way you found everything that year, on Twitter. At the time almost nobody had adopted it. But it was obviously the correct answer to a question we actually had: how does a person in Claude Desktop or Claude Code talk to your agent backend without you building a bespoke client for every surface? Betting on a standard early is one of the few asymmetric trades in engineering. It cost me days. Every MCP-speaking client that shipped afterward paid it back.

What happened

It never really worked. Users spawned agents; the agents mostly died. A couple dozen still run in production, flat, not growing. There is no framing of that as a successful iteration, so I won’t try.

And we were not uniquely bad at this. Within weeks of the January peak the agent-token sector lost nearly half its market capitalization;⁴ today the category trades more than 80% below it, with leading tokens down 90% from their highs.²⁵ By mid-2025, Gartner was predicting that over 40% of agentic AI projects would be canceled by the end of 2027, and estimated that of the thousands of vendors claiming to sell agents, only about 130 were real, the rest doing what they politely called “agent washing.”⁶ We weren’t the outlier. We were the median.

We weren't uniquely wrong. We were the median of a wave that was wrong together.

The anatomy of the failure

Strip the market drama away and the failure has a boring, reusable shape.

We started from the technology and searched for the user. Agents for creators, connected to paid API endpoints. Connected to what pain? A creator’s actual problems in 2024 were distribution and monetization, and an agent that could query CoinGecko didn’t touch either. When the origin story of a product is “the wave arrived,” the demand side is a hypothesis wearing a costume.

We were selling autonomy before the infrastructure for autonomy existed. The agents of that winter had no memory worth the name: session goldfish with a vector database bolted on. They had no safe place to act, so every capability was either neutered or terrifying. And until MCP, they had no standard way to be reached: every integration was bespoke plumbing. We built all three, badly-to-adequately, in-house, while also trying to find product-market fit on top of them. That’s two startups stacked in a trench coat, and the bottom one was the real one.

The failure list on the left cost the industry billions. The list on the right is just… true. It was true during the hype, it stayed true after the crash, and it’s the list I left with.

What survived

The part I keep coming back to: infrastructure conviction survives product failure. The agents died. Not one of the reasons they needed memory, isolation, and a protocol died with them.

So the lineage is direct, not poetic. The memory problem became LoomBrain, a second brain that lives on the internet precisely so agents can reach it while my machine sleeps. The sandboxing problem became sandcaster, agents as distributable packages running in isolated cloud sandboxes. The protocol bet became a habit: MCP went from “weird thing I found on Twitter” to the industry’s connective tissue, and being early to it once taught me what early feels like, which is mostly “slightly embarrassing, briefly.”

A hype wave is a terrible instrument for telling you what to ship. It is an excellent instrument for telling you what’s missing: you get to watch a thousand teams slam into the same three walls at once, and if you’re inside one of those teams, you get the wall’s texture at full resolution. The pivot never returned a product. It returned the missing list. I’m still working through it, and so far the list has been right about everything.