This article is based on Sentora's Report: The Smart Vault Paradigm. Download it here.
Most of DeFi was built with a human user in mind. The vaults, governance forums, and risk processes that run lending protocols today all assume there is a person on the other end.
That assumption is becoming the largest design flaw in the system. The rapid advancement in agentic systems will cause software to become one of the fastest-growing group of users in DeFi is: autonomous trading bots, AI-managed DAO treasuries, and agentic wallets that act on behalf of users without asking for input on every step.
These users behave very differently from human depositors and borrowers. They move faster, cannot read social media threads to decide whether a curation firm is trustworthy, and want highly specific risk profiles. The features that worked for human users, including weekly governance cycles, group-signed approvals, and reputation-based trust, are the same features these new users cannot work with.
The infrastructure underneath has to change to serve them.
Who (or What) the New Users Are
The shift is already happening across several categories.
DAOs increasingly use automated systems to manage their treasuries. Software allocates idle stablecoins, adjusts hedges, and moves capital between yield venues without a human approving each step. Trading firms run bots that open and close positions continuously, often hundreds of times a day. Agentic wallets let everyday users approve a goal, such as paying a bill or rebalancing a portfolio, and then carry out the underlying steps in the background.
These cases are becoming a meaningful share of total activity, and it’s reshaping market dynamics. Lending protocols that have not been designed for this kind of user will find an increasing portion of their volume going somewhere else.
Speed: The Bot Cannot Wait for a Committee
The first new requirement is speed. An automated trading bot that takes a position often needs to adjust its collateral within the same block. The decision to borrow more, repay, swap collateral, or close the position is part of one continuous loop the bot runs constantly. If the vault it borrows from operates on a different clock, the bot has a problem.
A human-curated vault, by design, cannot keep up. Its rebalancing happens on a schedule set by risk committee meetings, multisig approvals, and governance proposals. Even a fast cycle is measured in hours. A human depositor may find that reassuring, but a bot running thousands of decisions per day cannot use a counterparty that operates at that pace.
What it needs is a vault that updates on the same clock as itself. Anything slower introduces a delay the bot has to account for in every decision, which usually means routing the trade somewhere else.
Trust: Reputation Does Not Help a Machine
The second requirement is verifiable behaviour. A human user can extend trust to a curation firm based on track record, brand, audits, or word of mouth. An autonomous bot has no useful way to read those signals. The only thing it can check is the underlying logic of the system: the parameters, the code, and the rules the system has committed to. Human intent sits outside that domain.
This is why a multisig is a problem for these users. A multisig is a wallet controlled by a small group of people who must individually approve actions.
From a bot's perspective, the behaviour of that group is unpredictable:
Will the signers be available when needed?
Will they react quickly?
Are any of them compromised?
Automated systems have no way to measure any of this. It can only assume the worst case, which means it has to discount every interaction with such a vault.
What works for a bot is a vault whose behaviour is written into code and can be checked. If the rules are explicit and the system is committed to following them, the system can rely on the vault the same way it relies on the underlying lending protocol.
Reputation is a human concept that isn’t compatible with agentic systems. Verifiable rules are a different factor software can use.
Specificity: One-Size-Fits-All Does Not Work
The third requirement is precise customisation. Human users tend to accept generic risk profiles because designing a custom one is too costly. Automated users operate under different cost dynamics, and they want exposure shaped to fit their exact strategy.
A market-making bot, a leveraged staking strategy, and a DAO treasury allocator each have very different needs. Each one wants a vault with a specific collateral list, a specific price feed setup, specific liquidation rules, and specific liquidity guarantees. Producing vaults at this level of detail is too expensive for human teams. The fixed costs of risk committees and governance overhead do not scale to hundreds or thousands of specialised vaults.
Demand from automated users runs in the opposite direction. It pushes the number of distinct profiles up while pushing the budget per profile down. Meeting that demand requires an automated production model. Human-coordinated curation operates on cost structures that cannot reach the scale the demand implies.
When the Vault Becomes the Slow Layer
When most users are automated, a human-curated vault starts to look like a slow piece of middle layer software sitting between the bot and the underlying protocol. Every delay introduced by that layer, every uncertainty about how quickly a multisig will respond, and every limit on how specific a risk profile can be, has a cost. The bot either accepts less efficient capital deployment or routes around the vault entirely.
The way out of this is for the vault to behave like a piece of protocol. The current managed-product framing is what creates the delay. Once the vault operates at the same speed as the contracts beneath it, commits to behaviour that can be verified, and supports specific configurations at scale, the line between a vault and an agent starts to fade. Both are pieces of software with defined behaviour running on the same clock.
The Human Role: From Driver to Designer
Automated users change where people contribute in the system. The work shifts from execution to design.
A human curator today acts more like a driver. They watch the instruments, decide when to turn, sign each approval, and respond to events as they happen. That role scales poorly into an environment defined by speed and complexity.
A human designer sets the rules the system follows:
What kinds of collateral are acceptable?
What exposure limits apply?
What does the vault guarantee on withdrawals?
What is it trying to optimise for?
Once these rules are set and made verifiable, execution moves to the automated layer. The human contribution becomes the structure within which the system operates.
This is a more durable division of work for an environment where decisions need to be made many times per second. Humans bring judgement and policy while software brings execution.
Building for the Next Phase
As the user base increasingly becomes automated, the infrastructure must evolve to match their needs by providing speed at the block level and replacing reputation with verifiable behaviour. This transition requires support for highly specific risk profiles at scale and a clear division of labour where humans establish the policies that software carries out.
This is the direction Sentora is building toward. Deployment infrastructure that runs on the same clock as the protocols beneath it, with structured policies that automated users can verify and rely on, and a system designed for a world in which most counterparties on the other side of a transaction are themselves software.
The next wave of growth in DeFi will be driven by automated users, and the layer that serves them has to be designed for what they actually need.
The protocols that build for this transition early will absorb the volume that comes with it. Those that keep designing for a human-only user base will find themselves serving a smaller and smaller share of what the market actually does.
