In July 2024 I wrote a snapshot of AI and Web3 while those pieces were still mostly thesis. This is the update from the work since then: agents that transact, marketplaces for models and data, and payment rails that can sit inside a regulated frame.
I still ship at this intersection, and what changed in the year is that a few primitives now compose: programmable stablecoins, intent-based execution, non-custodial agent wallets, and off-chain inference with on-chain verification.
Stablecoins after the GENIUS Act
In July 2025 the United States enacted the GENIUS Act (Guiding and Establishing National Innovation for U.S. Stablecoins Act). The statute covers stablecoins and says nothing about AI, and it still changes how I design agent payments.
Its reserve rules are explicit, with one dollar of reserves per dollar of stablecoins issued, plus AML/KYC and issuer licensing. That does not make every AI and blockchain product legal, but it gives institutions a settlement asset they can hold without treating it as an unlicensed experiment.
Agents need that because they need a unit of account that does not swing while a tool loop is still running. In Agentic Finance I argued that agents need rails they can transact on under cryptographic constraints, and Google's AP2 and Coinbase's x402 are the payment standards I keep coming back to, since both carry intent, verification, and boundaries the user actually signed.
Stablecoins are the settlement layer under those protocols, and I covered that stack in Stablecoin Summer. The narrower point here is that once the onchain dollar is licensed infrastructure, an agent wallet becomes a product problem instead of a policy argument.
Agents as economic actors
Agentic finance, as I use the term, means an agent that can move capital inside limits you set, such as spending caps, approved contracts, time windows, and a wallet the user still owns, which is a different product from a chatbot that recommends a swap.
Three pieces are now real enough to build against:
Payment standards. AP2 and x402 give finance a shared shape for delegating intent, verifying execution, and keeping the cryptographic trail.
Non-custodial agent wallets. Platforms such as Crossmint's Agentic Finance issue wallets with programmable guardrails, so the agent transacts inside a fence the user defines.
Cross-chain coordination. Intent-based DeFi lets an agent say "swap 1 ETH for at least 3,000 USDC on the cheapest chain" and leave routing to a solver network, which is the constraint model agents need everywhere: declare the outcome and let someone else produce the calldata.
Adoption is uneven. A 2025 Wolters Kluwer survey put current agentic-AI use among finance leaders at 6%, with another 38% planning to adopt within a year, and Gartner has predicted that over 40% of agentic AI projects may be discontinued by 2027 on cost and unclear value. I read both numbers the same way, as rails that exist around an ROI that still has to be designed, and intents plus guardrails are how I keep an agent from becoming an unbounded spender.
Marketplaces for models, data, and compute
Most products still call a model through a centralized API. The other stack that matured this year is a set of networks that price models, datasets, and spare compute without a single vendor in the middle. I don't treat every token ticker as production-ready, but a few architectures are distinct enough to name.
Bittensor runs specialized subnets where models compete on a task, such as text, images, retrieval, or prediction, and get paid in TAO when the network judges their output useful, so quality is validated economically instead of by a curated model card.
Fetch.ai has deployed tens of thousands of autonomous economic agents (their published figure is over 23,000) for coordination problems like routing, energy, and DeFi automation, and the interesting part is agents negotiating inside an economic frame.
SingularityNET is the older marketplace shape: list an algorithm, let others consume it through a common API, and keep reputation and governance onchain.
Ocean Protocol is the data version, where providers keep the dataset and buyers run compute against it (compute-to-data) instead of downloading a copy. Medical, financial, and IoT data are the use cases it targets.
Morpheus is still the open-source personal-AI network I pointed at in 2024, with MOR rewarding people who build, host, and use a peer-to-peer agent network so inference stays owned by its users.
Around those sit compute and storage networks already running: NodeGoAI for unused capacity, OORT for a decentralized data cloud, and Akash, Render, Gensyn, and io.net on the GPU side. The pattern across all of them is that inference and training are expensive, idle hardware is everywhere, and a token is how the network meters both.
Data economics follow from that. A contributor binds examples to a wallet, a contract gates access, and the contributor gets paid when someone trains on them, while federated learning, homomorphic encryption, and zero-knowledge proofs let a model train without anyone handing over the raw file. When a product collects user data, that is the design I want, with the data in a vault the user controls.
How the primitives compose
The posts I have written this year describe layers of one stack: stablecoins as payment rails, intent-based execution so an agent does not have to know which bridge is healthy today, agentic finance for wallets and mandates, decentralized marketplaces for models, data, and GPUs, and Model Context Protocol for tools, memory, and logs that are not glued to one vendor SDK.
One flow I keep seeing in designs reads market data from a decentralized feed, expresses a cross-chain intent, lets UniswapX or Across compete to fill it, settles in a GENIUS-compliant stablecoin, and writes the receipt into MCP-shaped memory. None of those steps needs a custom bridge UI, and every one of them needs constraints the agent cannot silently widen.
That is how I approached product work at Bitlauncher too, where the AI tools and the token sale were parts of one piece of open infrastructure for AI startups.
What I change in the codebase
Agent-first UX means APIs that accept goals and constraints alongside button handlers, smart contracts that carry the fence, and signatures that carry the audit trail.
I still use hosted models when latency and quality win, and I don't hard-code a single vendor as the only inference path when a marketplace or self-hosted option is a real fallback.
If the product stores user data, I want revocation and, where it is honest, compensation, backed by deletion that actually happens.
MCP stays the boundary I want between model logic and business tools, and for money movement I default to stablecoin settlement, an intent protocol (CoW, UniswapX, Across), and a message layer (Wormhole, Axelar) instead of a one-off bridge.
What is still hard
Onchain inference is still slow and expensive, so the honest architecture is hybrid, with compute off-chain and verification onchain. Verifying inference without re-running the model is an open problem, and zero-knowledge proofs are the research bet there rather than my production default.
Marketplaces without curators attract junk and malice. Staking, slashing, and public benchmarks are the current answers, with Bittensor's validation loop as the clearest example, and none of them replaces knowing which subnet you called.
The GENIUS Act is narrow, and data privacy, model liability, and cross-border agent payments are still gray, so I build with privacy-preserving defaults and AML/KYC hooks I can turn on instead of assuming "decentralized" is a legal strategy.
Standards are proliferating, with AP2, x402, ERC-7683, and a pile of intent formats, and interoperability will come from the few that wallets and L2s actually implement.
Decentralized inference still costs more than a well-priced OpenAI call, and teams ship anyway with protocol subsidies and hybrid setups that keep inference centralized while training or verification is decentralized. I build inside those limits with named rails and named constraints, and I pick from the network catalog above only when one of them beats the hosted default for that product.