In an opinion column published on SiliconANGLE, Renee Davis, co-founder and chief business officer at OpenMatter Network Inc., writes that enterprise computing has long depended on chains of trust. Organizations have trusted cloud providers, software vendors, identity systems, and administrators to operate as designed. That model worked because humans remained the ultimate decision-makers. However, agentic artificial intelligence is changing that equation. Autonomous systems can retrieve information, make decisions, invoke external tools, collaborate with other agents, and execute actions on behalf of the enterprise. As organizations delegate more authority to intelligent systems, they also need a way to determine whether those systems behave as intended.
According to Davis, the question is no longer simply whether AI can perform complex work, but whether every consequential action can be independently examined and verified. The next era of enterprise computing will be defined not only by what autonomous systems can do, but by whether organizations can reconstruct and verify what they have done in practice.
The limits of traditional trust in information security
Davis explains that an AI agent may retrieve information from multiple enterprise systems, pass it to another agent, invoke an external service, and authorize a financial action. Although each individual component may be secure, the enterprise may still have no complete, independently verifiable record of how the final action occurred. Traditional security technologies can help establish who was authorized to act, what systems they could access, and whether unusual behavior occurred. Traditional logs can show that an application programming interface (API) was called or that a transaction took place, but these records do not necessarily capture the full chain of instructions, inputs, decisions, and actions that led to an outcome.
The central question, according to Davis, is whether trust itself is a sufficient architectural principle for a world of autonomous computing. Much of today's cybersecurity conversation begins with the question of which platforms to trust. While that is a reasonable question, Davis argues it is becoming less relevant. Traditional security technologies remain indispensable, but they were largely designed to protect systems through policy enforcement, behavior monitoring, and access control. They cannot, by themselves, establish which instructions and data shaped an agent's behavior, reconstruct its interactions across multiple systems, or demonstrate that the resulting record has not been altered.
Enterprise leaders should ask a different question when agents are involved: what evidence does the system produce, and can that evidence be independently verified? If the answer depends primarily on trusting the vendor, the infrastructure, or the software itself, the organization is still operating under the assumptions of the previous computing era. Autonomous computing requires systems that can produce independently verifiable evidence regarding what they were instructed to do, what information they used, and what actions they took. This is the difference between trusted computing and verifiable computing.
The shift from trust to proof
This transition does not require abandoning current cybersecurity practices. Identity management, endpoint protection, monitoring, and policy enforcement will continue to be indispensable. However, they become part of a broader architectural model in which independent verification adds a foundation for confidence. The first step is determining what behavior needs to be auditable. Not every interaction carries the same level of risk. Organizations should identify actions that could carry significant financial, operational, security, or regulatory consequences, and establish a higher standard of evidence for these activities.
For example, an agent summarizing an internal document does not require the same level of scrutiny as an agent approving a payment, modifying production code, or accessing regulated data. For each consequential action, organizations should define what the agent is permitted to do, what conditions it must satisfy before acting, and under what circumstances human approval is required. These requirements create a baseline against which the agent's actual behavior can later be evaluated.
This becomes especially important as agents begin acting across organizational and system boundaries. One agent's decision may depend on information generated by another agent, which may itself have relied on an external system. Without a verifiable chain connecting these events, organizations may know the outcome without being able to establish how it happened.
Documenting the action chain and preserving evidence
The next step presented by the author is determining what evidence is necessary to reconstruct these actions. This evidence should connect the original intent to the agent's behavior: what the agent was instructed or authorized to do, what information it accessed, what tools it invoked, what decisions or actions followed, and what outcome resulted. Depending on the use case, the audit log may also need to capture the initiating user or system, the policies and permissions in effect, communications with other agents, calls to external services, approval events, and any changes made to enterprise systems.
The objective is not to retain every piece of information an agent encounters, but to preserve enough evidence to reconstruct consequential behavior while respecting privacy, security, and data minimization requirements. Because agents rarely operate in isolation, evidence must be preserved across different systems. An audit trail that stops at the application level can miss critical parts of the chain. Every consequential workflow should have a consistent, traceable identity connecting the original request to subsequent delegations, tool calls, approvals, and outcomes. In addition, organizations should also determine what audit evidence they require from third-party agent platforms and external services so that accountability does not disappear when an action crosses a system or organizational boundary.
Cryptography's role in verifying evidence
The final question Davis raises is whether the evidence itself can be trusted. This is where cryptographic techniques can be powerful. Organizations can use digital signatures, cryptographic hashes, and other proofs to establish the integrity and provenance of records, inputs, and outputs. This makes it possible to demonstrate that evidence has not been altered and that a specific action is linked to the data and instructions associated with it.
Different mechanisms support different auditing needs:
- Digital signatures help establish a record's source.
- Hashes reveal whether instructions, inputs, outputs, or logs have been changed.
- Time-stamped attestations connect an action to a particular authorization, policy, or system state.
- Advanced cryptographic proofs can verify certain claims about an action without exposing all the underlying sensitive information.
However, Davis emphasizes that cryptography is only one part of the equation. A cryptographically protected record is not useful if the organization has not defined what to record in advance. Effective agent auditing requires clearly defined controls, comprehensive event capture, traceable connections between actions, and independent verification of the resulting evidence. Furthermore, cryptography cannot determine on its own whether an agent's judgment was appropriate or whether the data it received was accurate; its role is to establish the integrity and provenance of the evidence.
Operational decisions and periodic reconstruction tests
The appropriate level of verification should match the consequences of the action, rather than applying the same controls to every agent interaction. Davis recommends that organizations periodically attempt to reconstruct consequential agent actions using the evidence their systems produce. They should test whether they can determine what the agent was authorized to do, what information it relied on, how it interacted with other systems, and whether the resulting record remained unaltered.
The answers to these questions provide security teams with a way to investigate incidents, compliance teams with a way to substantiate decisions, and business leaders with a way to evaluate whether autonomous systems are operating within established boundaries. Building this capability requires making operational decisions prior to deploying agents. These decisions include: which actions require enhanced auditing, what evidence must be captured, how long it should be retained, how records will be connected across systems, who will have access to them, and who is responsible for reviewing them. Organizations should also establish a recurring process to test the ability to reconstruct consequential actions and correct any gaps that are revealed. The broader shift is from designing systems that are merely trusted to systems whose consequential behavior can be examined and proven.