What 2025 actually was

Two things happened in 2025. The technology actually shipped. The casino emptied out. They happened simultaneously, which is the most 2025 thing possible.

the agents part

MCP went from a curiosity to infrastructure. I watched people dismiss it in January and integrate it into their whole stack by October. The Linux Foundation took it on in November, which is the institutional signal that says “this is plumbing now, not a demo.” When a protocol gets donated to a foundation, it means the people who built it decided the right move was to let it be boring infrastructure forever rather than a product. That’s the compliment.

Claude Code went generally available in May and changed how I work more than anything else this year. Not in the “AI writes my code” sense. More like the “I can stay in the editor for six hours without reaching for a browser” sense. The agentic-coding tools matured past toy status. GPT-5 dropped in August, Claude Opus 4.5 in November, Gemini 3 Pro in November. Every few weeks there was a new frontier model and the benchmarks kept moving. I stopped tracking individual scores sometime in the fall because the benchmarks themselves were getting gamed. SWE-bench Pro exposed that in September and it kind of broke the narrative that any single number meant something.

x402 is the one I didn’t see coming. Coinbase opened it in May as a protocol for agent-to-agent payments and it sat quiet for a few months. Then the x402 Foundation launched in September and usage went vertical in mid-October. The idea is simple: agents can pay agents, without a human in the loop, using crypto rails. It sounds like a crypto-twitter fantasy but it’s running in production now. That’s a different kind of real.

The honest summary on AI: the infrastructure year. Models got better, tooling got real, the agentic narrative graduated from hype to plumbing. That’s good, even if it’s less exciting to write about than “AGI this quarter.”

the market part

October 6. BTC printed ~$126K. All-time high, everything looked good, the cycle was supposed to keep going.

October 7. About $19B liquidated in a day.

What followed was a slow, grinding walk down. By November the “bear market” framing was everywhere. December has been heavy. SOL sitting in the low hundreds. Sentiment in the basement. The people who were loudest in the summer have mostly gone quiet or found something else to be loud about. This is how it always goes and it still surprises everyone every time.

The pump.fun creator-rewards dynamic that made “build in public with a token attached” financially viable for most of 2025 got harder in the back half of the year. The mechanism still works but the market being cooked means the rewards are smaller and the holders are angrier. You build a thing, you ship a commit, and someone tells you the commit isn’t good enough because the chart is down. The chart being down has nothing to do with the commit, but the correlation feels real to someone checking the price every twenty minutes, so.

what I actually did

Launched SAM in September. Open-source agent framework for Solana: tools, a keyring, risk checks, real chain integration. Also launched a token the same day, on the same wallet I’d used before, and got sniped in the first minute. I wrote about that already. The framework was solid. The launch mechanics were embarrassing. Both stayed true.

October was integration month. Hyperliquid, Aster for perps. Multi-agent support. x402 payment tools wired in. Kalshi and Polymarket for prediction markets. The “AI-Fi economy” framing started making sense around then because you could actually demonstrate it: an agent, a wallet, real on-chain actions, plus the ability to pay for data or execution through a protocol instead of an API key. That worked.

November and December were different. Bear sets in, you stop getting dopamine from the chart and you have to decide whether you’re building for the technology or for the number. The unsexy holder front-end got built. I got fried. I kept going anyway, which is either discipline or stubbornness. The line between those is a matter of outcome.

The year I shipped a real thing into a real chain with real stakes, then watched the stakes get colder. The tech held. The market is the market.

That’s 2025. The tools are better than they’ve ever been. The infrastructure is real. The casino has a bad hangover. Somewhere in that gap is where the actual work lives, and that’s where I’ll be.