Agentcoin - Co-Founder

When large language models went mainstream, Polywrap’s central bet—cross-language client SDKs—stopped matching where teams actually felt pain; large language models and agents were clearly the future of how humans build and interact with software. We pivoted and continued under the name 'Agentcoin' to work where the industry was clearly heading next.

Vision: Why “Agentcoin”

Agentcoin's north star was to become a network of specialized agents on decentralized protocols, built to generate real revenue and align incentives onchain so useful AI agents could share upside with humans instead of quietly displacing them. We described it as symbiosis: open networks, human judgment for values, agents for adaptive execution.

How We Operated After the Pivot

We shrank the team to move faster and shifted away from deep infrastructure toward end-user products with clear audiences—after roughly two years on a developer toolchain, we wanted tangible use cases we could name, demo, and market. Each cycle was a new, tightly scoped product every month to month and a half, ending in something we could put in front of real users. LLM capabilities, agent UX patterns, and the competitive field were all moving too quickly for a years-out fixed plan; so instead we aimed to ride the wave of innovation through tight product release cycles.

Evo.Ninja — AutoGPT Arena Hackathon Winner

We built Evo.Ninja to win the AutoGPT Arena hackathon: at the time, AutoGPT was the flagship open-source “generalist agent” project on GitHub, and the contest ranked entrants on a demanding benchmark suite—the winner earned the title of best generalist agent and prominent placement alongside that repo. We competed to top those benchmarks and did, earning the Current Best Agent: evo.ninja spot on the leaderboard.

We then shipped Evo.Ninja as its own product—an agentic chat that could plan, call tools (research, spreadsheets, coding), upload documents, and run generated code inside a workspace tied to the conversation, at a time when ChatGPT still lacked much of that end-to-end loop for mainstream users. On the order of ten thousand people used it. Once OpenAI shipped native code execution, file uploads, and customizable agents, competing with ChatGPT as a small team was unrealistic; we stepped back from Evo.Ninja as a generalist consumer surface and redirected effort toward narrower, Web3-native agent products where we could differentiate.

Web3 Agent Use Cases — Rapid Product Releases

For the next wave of work we narrowed to tailored Web3 agents because the category still had a lot of untapped, user-shaped problems—funding, trading, social feeds, quests—where the default experience was expert tooling and long UIs, and we believed agents could let people state intent in natural language and get guided execution instead of memorizing flows. We also assumed the largest AI platforms would not prioritize crypto-native journeys, which gave us room to ship focused products. Under the Agentcoin banner we shipped a series of agents, each with a notable industry partner and a clear use case:

• FundPublicGoods.ai (with Gitcoin): a collaboration that used the full corpus of Gitcoin projects so users could describe what they care about (for example, 'climate' or 'cryptography') and get back a report of candidate projects, an impact-style grading pass so funders could see how the system scored alignment and evidence of impact, and clear paths to split funds to onchain addresses where supported. It was the first of our web3 products where people kept saying the flow felt 'magically' easy compared to the manual alternative.
• AutoTX (with Biconomy): a transaction and trading agent that did market research in natural language (for example, 'blue-chip DeFi tokens' or 'top movers in the last 24 hours'), proposed allocations, and used a smart account the user funded so the agent could swap and rebalance positions—an early 'tell the agent what you want, it plans and moves funds' experience.
• Prediction Prophet (with Gnosis AI and Olas): a forecasting agent for prediction markets—deep web research plus statistical legwork in service of more informed positions. It built on the Evo.Ninja-era web research stack; a version of the agent remains in operation on the Olas network today.
• Indexer (Farcaster, pre-dating Grok): a social agent on a decentralized, X-like social graph. Because Farcaster is an open protocol, we could work from the public firehose and let people mention @indexer on the feed to ask about a post, request recommendations ('best cat pictures right now'), or get a 'what's going on here?' summary—an early parallel to the later assistant-on-the-timeline pattern people now associate with Grok on X, but in a protocol-native, credibly neutral form.
• SuperAgent (with Optimism, for Superfest): a quest-completion agent for Superfest, Optimism's multi-chain quest campaign. It helped users work through the quest graph with routing that depended on their balances and which quests they still had open—mixing transaction planning, chain selection, and progress tracking so people could move through the program without drowning in manual steps.

The Lessons Learned and Our Evolution into AITV.GG

Across Evo.Ninja and the Web3 agents above, the honest read was the same: we could earn spikes of attention, but almost nothing held month-over-month growth once the novelty wore off. Utility-shaped agents turned into a crowded aisle—trading, prediction, planning, and research copilots all sounded interchangeable on a timeline—so “we shipped another capable agent” was rarely enough to win distribution against incumbents and fast followers. The signal that kept repeating was different: people gravitated to agent experiences that felt like entertainment—character-led, surprising, easy to share—and we kept circling a whitespace personal assistants ignore: multiplayer agents meant to be used and owned collectively on the open internet, with Web3 as the coordination substrate. Those lessons are what ended the Agentcoin chapter as a utility sprint and became AITV.GG: live, multiplayer agent entertainment where the product is the show, not a marginally better sidebar tool.