Web3 developers are leaving crypto, and the work for those who stay is moving toward AI. Electric Capital counted 25,623 monthly active crypto developers in late October 2026, down 36% in a year; nobody has measured where the leavers went. What can be measured is what the stayers are asked to build: at ETHGlobal's in-person hackathons, sponsors asking for AI or agent work held at about five per event from 2025 to 2026 while every other kind of sponsor fell from about 21 per event to 5, so agent tracks went from 18% of sponsors to 53%. The Web3 skill that carries over is adversarial engineering, and the evidence a reviewer can check is the same in both fields.
Facts and counts checked on 5 November 2026.
Is the crypto developer pool shrinking?
Yes. Electric Capital's developer dashboard put monthly active crypto developers at 25,623 in the week of 28 October 2026, down 36% in a year and well below a peak of 45,931 in June 2025. The experienced core is shrinking too: developers with two or more years in crypto fell 25% over the year. They are still 11% above three years ago, likely because developers who arrived two and three years ago have now been around long enough to count as established.
Some of those who leave take their specialty with them. Sam Blackshear, co-founder and CTO of Mysten Labs, the company behind Sui, announced in August 2026 that he was joining Anthropic to do defensive security research. His turning point, he said at a security panel he hosted in April, was using Claude to port a static analysis tool he had built at Facebook over to Move, the language Sui contracts are written in.
How many Web3 hackathon sponsors now ask for AI agents?
At ETHGlobal's five in-person events in 2026, 29 of 55 sponsors put AI or agents in a track title, from a third of sponsors in Mumbai to 10 of 16 in New York. Sponsors fund the integrations they want built, so a prize page is a fair read of what Web3 companies want from developers, and the same count on the 2025 pages gives the baseline.
AI and agent tracks at ETHGlobal in-person events, 2025 and 2026
| Event | Dates | Sponsors | Sponsors with an AI or agent track title | One such track |
|---|---|---|---|---|
| Taipei | 4 to 6 Apr 2025 | 18 | 3 | Celo, "Best AI Agent on Celo" |
| Prague | 30 May to 1 Jun 2025 | 14 | 1 | Hedera, "AI, Agents & Hedera Services" |
| Cannes | 4 to 6 Jul 2025 | 28 | 4 | Artificial Superintelligence Alliance, "Innovator's Edge: for most impactful and useful AI Agent" |
| New York | 15 to 17 Aug 2025 | 27 | 8 | OpenSea, "Best AI Experience with OpenSea MCP" |
| New Delhi | 26 to 28 Sep 2025 | 21 | 4 | Polygon, "Best Use of x402 Agentic Payments on Polygon" |
| Buenos Aires | 21 to 23 Nov 2025 | 46 | 8 | Circle, "Best AI Agent Wallet with Circle Wallets" |
| Cannes | 3 to 5 Apr 2026 | 12 | 6 | Ledger, "AI Agents x Ledger" |
| New York | 12 to 14 Jun 2026 | 16 | 10 | Google Cloud, "Best On-Chain Agent Economy Application" |
| Lisbon | 24 to 26 Jul 2026 | 8 | 5 | The Graph, "Best AI Tooling for The Graph" |
| Tokyo | 25 to 27 Sep 2026 | 7 | 4 | Curvegrid, "Best AI Agent Project" |
| Mumbai | 5 to 7 Nov 2026 | 12 | 4 | 0G, "Physical AI: Agents with a Body" |
Compiled from the prize page of each in-person hackathon on ETHGlobal's event calendar on 5 November 2026: 28 of 154 sponsors (18%) in 2025 against 29 of 55 (53%) in 2026. A sponsor counts when at least one track title names AI, an AI agent, "agentic", an LLM or an agent product such as World's AgentKit; ETHGlobal's own prizes and the finalist perks are left out, and without World's AgentKit tracks New York 2026 would be 9 of 16. Online events, including the agent-only Agentic Ethereum (2025) and Open Agents (2026), are excluded, and Mumbai was still running on the day of the count.
Most of that jump is sponsors leaving, not sponsors converting. Per event, AI and agent sponsors went from about 4.7 to 5.8, while everyone else went from about 21 to 5. Agent work is the part of the sponsor budget that survived the downturn, and it is specific: Web3 companies want agents to use their rails (wallets, payment flows, name services).
Several finalists took that brief seriously. By our reading of their project pages, six of the ten ETHGlobal Cannes 2026 finalists were agent projects (ENShell, DIVE, Corpus, ALMA, maki and npmguard), and half of those were built around limiting what the agent can do. ENShell runs every action "through a four-layer security pipeline before a single wei moves"; ALMA enforces an onchain whitelist of allowed calls with a revocable 30-day key; maki keeps key material in hardware, and its page says the model "interprets intent but never moves funds".
That is the same instinct Vitalik Buterin set out in April 2026 in his post on running LLMs securely: require human plus LLM 2-of-2 confirmation for risky actions, and allow a daily limit of about $100 to bypass the human. Smart-contract developers have spent years assuming every input is hostile and every key is a target, and an agent that holds a wallet needs exactly that habit.
The standards those tracks named point the same way. ERC-8004, a draft Ethereum standard from August 2025, gives agents identity, reputation and validation registries, and x402, an open payment standard documented by Coinbase's developer platform, lets a service charge a request whether "the buyer is a human or a machine". Google Cloud's 2026 New York track asked for both.
Is AI work a Web3 thing or an everyone thing?
An everyone thing. A count of the 678 top-level posts in Hacker News' October 2026 "Who wants to be hired?" thread, using a broad keyword list, finds that 461 of them, 68%, mention AI. Of the 34 who mention Web3, 26 do (76%), a little above the thread's rate. Web3 developers are doing what every developer on the market now does, and only about 1 in 20 job-seekers mention Web3 at all. A post counts even when it says "no AI work", so this measures what people put in front of an employer, not what they can do.
The job market shows the same spread. In the Indeed Hiring Lab's AI tracker data, 7.34% of US job postings mentioned AI at the end of September 2026, up from 3.71% a year earlier.
Are AI tools turning specialists into generalists?
They widen what one builder can attempt, and the review, testing and deployment work does not shrink with it. No dataset yet measures specialists becoming generalists, so the answer is pieced together from studies of parts of it.
METR's 2025 randomized study found 16 experienced open-source developers took 19% longer with AI tools on their own repositories, while expecting to be 24% faster. Its February 2026 update, with newer tools and 57 developers, estimates a speedup but calls its own signal unreliable. For smart contracts, Diana Levytska of AstraKode wrote in April 2026 that "generating code was never the real bottleneck" (her post on where the bottleneck moved): the bottleneck was testing and safe deployment.
The tools have adjusted to that bottleneck. Since February 2026, BuidlGuidl's Scaffold-ETH 2 has shipped an AGENTS.md file that coding agents read for project context, plus skill files, from ERC-20 to Solidity security, that an agent loads before it writes code.
What do reviewers trust when everyone lists AI?
Work that someone other than the builder stands behind, because the usual signals stopped working.
Take-homes went first. Tristan Hume, a lead on Anthropic's performance optimization team, wrote in January 2026 that on the company's own take-home, "Claude Opus 4 outperformed most human applicants" within the time limit (Anthropic's post on AI-resistant evaluations). Its replacement simulates novel work, and judgment about how to invest in tooling is part of the signal. In a Karat survey of 400 engineering leaders published the same month, 71% said AI makes technical skills harder to assess.
Skill badges answer a narrower question. LinkedIn's verified AI skills, launched on 26 January 2026, are issued by tool vendors such as Replit and Lovable based on how you use their product. That shows you used the tool, not what you built with the tool or whether it held up.
Simon Willison's December 2025 standard for code review is the useful frame: "Your job is to deliver code you have proven to work" (his post on code proven to work). Stretch it from one pull request to a career and the evidence that a stranger in either field can check looks like this:
- a pull request a maintainer reviewed and merged in a repository you don't control, where the thread shows a real change
- a placement the organizer names on its own results page (a finalist spot or a top-three track finish, not a share of a pool prize), with your team
- a role your employer confirms, rather than one you typed
- a finding a third party paid for or ranked: an audit-contest result on Code4rena, Sherlock or Cantina, a bounty paid on Immunefi, or an eval an outside party ran
A demo with tests is still worth having, but it is your claim until someone else runs it.
These live in different places, and a reviewer rarely chases more than one. A Dapping profile puts the projects you shipped, the roles verified companies confirmed and the wins organizers recorded on one proof page, so a reviewer sees what you built and who signed off on it in one link.
Start with the agent work you would most want a reviewer to see. If it has none of the four behind it yet, the cheapest one to get this month is usually a maintainer's review on an upstream fix, or an entry in the next agent track at a hackathon that publishes its results.
Why we widened Dapping to builders in Web3 and AI
In October 2026 we widened our line from Web3 to Web3 and AI: Dapping is now the verified professional graph for builders in Web3 and AI. We made the call in mid-October on earlier readings of the same sources, Electric Capital's decline, the agent tracks at ETHGlobal New York and how developers describe themselves when they look for work. The fuller counts compiled for this piece point the same way. A line that named half of a builder's work was describing less of it every month.
A verified professional graph is a network of people, companies and work in which every claim shows whether someone else stands behind it. Nothing in that depends on a chain. When a company Dapping has verified confirms a role, the profile marks it, so a reviewer can see which roles are backed. The methodology page explains how to read each kind of proof.
What changed is what sits around builders. AI companies now have their own categories (AI x Crypto, AI Applications, AI Infrastructure and AI Research), and their jobs, hackathons and programs are being added the way the Web3 catalog was built. An AI project you add sits on the same proof page as your Web3 work and your confirmed roles.
Whatever you call the field on your next application, a reviewer will ask the same question of every line on it: who else stands behind this? Answer it before they ask: list the agent work you most want reviewed as a Project with your teammates credited, next to the roles that already have someone behind them.
Frequently asked questions
Should I call myself a Web3 developer or an AI engineer?
Name the work instead of the field. When about two-thirds of job-seekers on Hacker News mention AI, the word separates nobody, and "Web3 developer" undersells an engineer who builds agent wallets. A headline such as "smart-contract and agent engineer, built a spend-limit layer for agent wallets" tells a reviewer both stacks at once, and the link under it should open the backed version of that claim.
What counts as proof for a pseudonymous builder?
The same outside parties, tied to the handle instead of a legal name. Audit-contest results under a consistent handle, organizer-named hackathon placements, bounties paid to the handle's address and merged pull requests all attach to a pseudonym. A contract deployed from an ENS-named address shows the handle shipped something; it becomes proof once an auditor or an integrating protocol stands behind it. The one item that usually needs a real identity is an employer-confirmed role.
Are Electric Capital's and Artemis's developer counts comparable?
No. Electric Capital counts monthly active developers who author original code in crypto repositories. Artemis, as reported by CoinDesk in March 2026, counted about 4,600 active developers, down 56%. Both point down. Quote each with its own method, and never credit the Artemis decline to Electric Capital.
Top comments (0)