
Written by: Christian, a16z Frontier
Translated by: AididiaoJP, Foresight News
AI Agent's Implementation Hindered by Permissions: Opportunities for Crypto Teams
Our crypto team has had many discussions with AI infrastructure companies, initially thinking that payments would be the biggest entry point. Stablecoin settlements are faster, cheaper, and easier to manage, which seems reasonable.
However, actual conversations have not reflected this. Compute and inference providers are more concerned with growth and fundamentals: how to secure capacity, how to get deployments running, and how to maintain gross profits. Teams creating Agents repeatedly emphasized reliability and control—what the Agent can interact with, which actions must be approved first, and how the company can confirm what actually occurred.
Payments remain a long-term direction for crypto and AI, but the clearer current need is verifiability. As enterprises delegate more work to Agents, they must be clear about: who approved the operation, what the permission boundaries are, and whether the authorization is still valid. Crypto teams have long been involved in signatures and delegated authorizations, which can seamlessly connect to this area.
Define Boundaries Clearly to Safely Delegate Authority
Delegating tasks to Agents requires them to be reliable while also clearly limiting their permissions. A permission system does not resolve the issue of Agents consistently making mistakes; however, once an Agent performs well and is deemed useful, companies still need to decide: which systems it can access, which actions must be approved, and how to revoke permissions at any time.
This is especially important in new systems. Whether the Agent is helping users adjust company information or aiding developers in modifying production systems, companies must first define boundaries and ensure that they can enforce them.
For example: a company uses an external software Agent to investigate production accidents. The Agent can read logs and propose fixes, but it must first gain approval before modifying the system. The company authorizes, the provider runs the Agent, and the receiving system checks permissions—all of this needs to be verifiable. A survey of 150 IT and security decision-makers in January 2026 revealed that 69% believe security concerns are slowing down the implementation of Agents, while only 27% believe that existing identity systems can effectively manage Agents on a large scale. Data leaks and excessive permissions are primary concerns.
Research on deployed Agents points to the same issues. In "Measuring Agents in Production," reliability is cited as the biggest development challenge. Teams often limit Agents to short processes and retain human intervention. Jellyfish's engineering data also shows that while Agent involvement is increasing, it is still rare to see demands lead all the way to a pull request in a fully automated process without human involvement.
Encoding Agents is another example: they can produce useful changes, but each deployment still requires human approval. Companies must clarify which systems they can modify, which actions need review, and how to revoke permissions. Better control is essential for companies to feel comfortable delegating more work.
Cross-Company Collaboration Requires a Trusted Record for All Parties
When Agents operate across companies, the issues become more complex. The company authorizes, the provider runs the Agent, and the receiving system enforces compliance—all three parties need a common basis to confirm "was this operation really authorized?"
Existing identity systems and application permissions already cover some of the needs. The opportunity for cryptography lies in making authorization, execution, revocation, and auditing across systems simpler. The crypto team's experience in signing, delegating permissions, and user-owned accounts can be effectively utilized here.
Existing cases illustrate this point. In the Turnkey Wayfinder case, Parallel Studios used session-based signatures, allowing users to authorize multiple transactions with a single authorization while still retaining control over their wallets. Wayfinder itself focuses on on-chain transactions, but it is more noteworthy that it allows software to act without granting it unlimited permissions. A bigger opportunity is to enable Agents to perform more tasks within a range of permissions that can be executed and revoked by the enterprise, leaving a verifiable record behind.
The value of blockchain lies in the need for multiple parties to share a permission record that no one can control alone, or the assets themselves are already on-chain. However, all parties must first agree on "who can authorize, what permissions explicitly allow, and how to revoke"—this is precisely what blockchains excel at.
For off-chain applications, companies can place Agent permissions, expiration times, and revocation records on-chain. This registry runs parallel to existing off-chain identity and application systems, allowing the receiving system to verify signature requests before releasing operations. Both the company and the provider can check authorization content, with the receiving system responsible for enforcement.
Most Agents currently operate off-chain. To reach enterprise customers, built-in controls must integrate into their existing systems. For crypto teams, this means extending products beyond pure on-chain scenarios. There is only one key test for market entry: can target customers delegate tasks that were previously stuck due to permissions or approvals to the Agent?
Start Selling to Enterprises with Specific Stuck Tasks
Telling enterprises directly, "you need to trust your Agent" makes it hard to close a sale. A better approach is to focus on specific tasks they want to delegate but are hindered by permissions or approval processes, demonstrating how your product can remove barriers and whether this improvement is worth introducing a new system—whether it's worth it, and if it can improve profits or open new business opportunities.
Buyers depend on the context.
When deploying Agents internally, decision-makers are usually the identity or security engineering leads, with the CISO potentially being a sponsor. Pain points commonly include excessive permissions, slow revocation, or difficulties reconstructing actions during cross-system investigations. During sales, focus on a specific deployable instance that can be unlocked and explain how it can integrate into the existing security framework.
When selling Agent software to enterprises, approach it through the CTO or the engineering lead for the Agent platform. First, inquire how much custom authorization each client deployment requires, whether it slows down the roll-out or consumes engineering resources, then showcase how your product can meet requirements with less custom work, speeding up deployment.
In either scenario, founders need to clarify: who maintains the current temporary solution, what the costs are, and whether they are more likely to resolve it themselves or willing to pay vendors.
Existing identity vendors have a first-mover advantage because their systems are already in operation. Even if new products offer better control, clients will weigh integration and maintenance costs—some salespeople describe this as battling gravity. You must clearly understand the shortcomings of existing systems and why clients would choose you over extending old systems. The threshold isn't just filling gaps; improvement must significantly exceed the cost of adoption. Unlocking important Agent rollouts might warrant significant integration efforts, while small operational improvements must be exceptionally easy to implement.
Begin with very specific workflows, ideally taking on the integration work yourself without throwing the entire burden onto the client. If discussions confirm an opportunity, conduct a narrow pilot: allow the Agent to complete specific actions under permissions the client can execute and revoke, while relaying records back to existing auditing tools. Test whether this reduces integration work or allows clients to delegate tasks requiring manual review, while also blocking unauthorized actions.
When evaluating client opportunities, clarify the following questions before discussing:
- What tasks does the client want to delegate? What is preventing that today?
- Who is maintaining the temporary solution? What are the costs?
- What kind of improvements are worth introducing a new system?
- What does adoption require from the client? How much can you take on?
- What does your solution make possible or significantly easier that existing tools cannot achieve?
These questions should be clear before contacting clients, not afterward.
Payments Remain Important, but Should Come After Control
New payment infrastructures will only be adopted when they remove friction for what customers truly want to accomplish. The growth of Agent adoption alone cannot tell you which processes need new payment solutions. Specific needs depend on the context:
- If an Agent needs to access unfamiliar APIs, immediate payment may be necessary;
- If a coding Agent utilizes existing enterprise subscriptions, there may be no payment friction at all;
- Compute providers may benefit from different settlement methods, but urgent operational issues often take precedence.
Improving payment becomes a priority only when current settings block customers from doing what they want: needing to manually set up accounts to purchase services, billing models making small purchases unfeasible, high operational costs tied to settlements—these are the reasons for adopting new systems.
Payments also require control. Take AP2 as an example; it uses cryptographically verifiable authorization to record what users permit the Agent to do (including purchases when users are not present). However, clients may only need better control without necessarily requiring a new payment method. An increase in transaction volume alone cannot prove the need for new payment infrastructure unless you understand what the client is buying and why existing options fall short.
Conclusion
Before entering the Agent economy, the real question is not how to bring stablecoins into AI operations but rather what can make these companies comfortable delegating more responsibilities to Agents. Crypto and other frontier teams should spend time collaborating with those filling this gap to solve problems. This is just like selling anything new to enterprises: you must start from where the customer currently is, not from where you hope they might go.
免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。