AI agent demos proliferate
Several AI agent demos appeared: a Kaspa KANet demo showed agents executing cross‑chain sells, a Bittensor subnet reportedly outperformed a 4B model after two weeks of training, and HeyGen published a CLI for generating and shipping AI video agents. Each release was presented as a proof‑of‑concept for on‑chain or near‑chain agent activity. (x.com) (x.com) (x.com)
Artificial intelligence agents are moving from chat windows into software that can place trades, train models, and ship videos with a command line prompt. Three demos posted in early April showed that push across crypto and media tools. (heygen.com) (docs.heygen.com) (kaspa.org) An artificial intelligence agent is software that takes a goal, decides on steps, and calls tools without waiting for a human after every click. In these demos, the tools ranged from a blockchain transaction flow to a decentralized training network to a video generation interface. (heygen.com) (docs.heygen.com) (taostats.io) HeyGen’s public documentation says its Video Agent can generate a finished video from a single text prompt through an application programming interface endpoint, and its Model Context Protocol server connects that capability to Claude Code, Cursor, Gemini Command Line Interface, and other agent software. HeyGen says the Model Context Protocol version uses OAuth instead of exposed application programming interface keys and is available on all plans. (docs.heygen.com) (heygen.com) Kaspa is a layer one blockchain built on a block directed acyclic graph, a ledger design that allows parallel block creation instead of a single-file chain. A demo tied to Kaspa’s KANet project was presented as an example of an agent executing a cross-chain sell, meaning software initiated a trade that moved value between networks rather than only answering a question about markets. (kaspa.org) Bittensor organizes its network into subnets, which are separate markets where participants compete on specialized machine learning tasks. The network’s public block explorer listed 47 subnets on April 14, 2026, with live data on stake, emissions, and trading activity. (taostats.io) Recent reporting on Bittensor’s Subnet 3 said the team trained a 4 billion parameter LLaMA model in a distributed setup, and later claims on social media said a subnet beat a 4 billion parameter baseline after about two weeks of training. Public documentation available through search confirms the 4 billion parameter training milestone, but not the benchmark details behind the “outperformed” claim. (phemex.com) (taostats.io) Those releases arrived as companies and crypto projects tried to make agents do work near a blockchain or through external tools, not just generate text. HeyGen’s product pages describe video generation, retrieval, and management as agent actions, while the Kaspa and Bittensor demos framed trading and model training as agent tasks that can run with limited human intervention. (heygen.com) (docs.heygen.com) (kaspa.org) The common thread is not that these systems are fully autonomous today, but that each demo packaged one concrete action into an agent workflow: make a trade, train a model, or publish a video. The next test is whether those proof-of-concept clips turn into products with repeatable results, public benchmarks, and users beyond the demo accounts. (docs.heygen.com) (heygen.com) (taostats.io)
