Agents Under the Hood: Building the Agentic Stack

About this session

Five panelists agreed a basic call-the-model loop does not make an agentic system, and layered in memory, durable data storage and enough autonomy to act without constant human sign-off, but split over how much of that stack needs to be decentralised. Nicolás Montone argued agents mainly need to trust whichever model provider sits behind them, the same trust any app already asks for, and that decentralisation earns its place mostly around agent-to-agent payments, where he pointed to x402 as the protocol drawing the most investment. Nader Dabit and Juan Irungaray leaned the other way, expecting agents built by separate teams to eventually need something resembling blockchain-level trust between them.

Montone broke his own stack into four layers: an environment layer where his team ran agents on Vercel's fluid compute, serverless billing that only charges for CPU time actually spent rather than time waiting on a model; a context layer of cacheable, injectable system prompts; an actions layer of tools and MCP; and an application layer where a failed build gets fed back to the agent so it can self-correct.

Dabit argued a smart contract cannot host an agent, since contracts are action-based while agents run continuously and call external APIs. His answer was Eigen's own compute product, a trusted execution environment that generates a wallet key inside the runtime itself, so funds sent to an agent can only be spent by the running agent, not its operator. He described a companion product built for deterministic, re-executable inference, so a given prompt and model could be shown to reproduce an output, and called ERC-8004 useful for reputation.

Clemens Wan focused on limits: an agent with wallet or RPC access needs hard caps on trade size, plus an abstraction like a phone user never picking a GPS satellite, just setting a preference for cheaper, faster or greener. He wanted approval flows that check with a human until a workflow proves itself and gets remembered, rather than an intrusive prompt every time, alongside attested identity so an agent can prove who it acts for.

Irungaray described Google's Agent Engine as infrastructure for hosting agents built on LangGraph, Hugging Face or Google's own ADK, handling memory and session management for developers. He placed ADK's level of abstraction between LangGraph and n8n.