The race to build autonomous AI agents that can operate over extended timelines is heating up, but according to Rayan Garg, a key voice in the startup ecosystem, most of these systems are missing a critical piece: robust verifiers. In a recent discussion highlighted by StartupHub.ai, Garg argues that as AI agents take on longer and more complex tasks, the need for reliable verification mechanisms becomes not just helpful, but essential. Without better verifiers, these agents risk making compounding errors that could undermine trust in AI-driven automation.

The Growing Complexity of AI Agents

AI agents are no longer confined to simple, single-step tasks like answering queries or scheduling meetings. Today’s agents are being designed to handle multi-step workflows that can span hours, days, or even weeks. These long-horizon operations might involve everything from managing a supply chain to conducting deep research or executing financial strategies. As the scope expands, so does the margin for error.

Garg points out that the current verification tools were largely built for short-term interactions, where a human can quickly spot and correct mistakes. But in a long-horizon setting, the agent operates with far less supervision. Each decision builds on the previous one, meaning a small error early on can snowball into a major failure later. This makes the role of verifiers — systems that check the agent’s outputs and decisions — far more consequential than in traditional AI setups.

What Exactly Are Verifiers?

In the context of AI, verifiers are mechanisms that assess whether an agent’s output is correct, safe, or aligned with user intent. They can be rule-based, model-based, or even human-in-the-loop. For short tasks, a simple check may suffice. But for long-horizon tasks, verifiers need to evaluate intermediate steps, not just the final result. Garg emphasizes that this is where current systems fall short.

Why Current Verifiers Aren’t Enough

The core problem, as Garg sees it, is that most verifiers today are designed for short-horizon tasks. They assume a single input-output pair and a clear ground truth. In long-horizon scenarios, however, there is no single “right answer” at each step. The agent must make judgment calls, adapt to new information, and sometimes take detours to achieve the final goal. A verifier that only checks the final output cannot catch subtle missteps along the way.

Moreover, long-horizon agents often operate in dynamic environments where conditions change. A verifier that was accurate at the start of a task may become outdated by the end. This requires verifiers to be adaptive, continuously updating their criteria based on new context. Garg argues that building such adaptive verifiers is a major technical challenge that the industry has yet to fully address.

  • Error compounding: Small mistakes amplify over long tasks.
  • Lack of intermediate checks: Most verifiers only assess final outputs.
  • Dynamic environments: Static verification criteria become obsolete quickly.
  • Scalability issues: Verifying each step in a long process is computationally expensive.

The Path Forward: Better Verification Strategies

So what can be done? Garg suggests that the solution lies in developing multi-layered verification systems that combine several approaches. One promising direction is the use of process-based verifiers, which evaluate the reasoning steps an agent takes, rather than just the outcome. This is similar to how a math teacher checks the work, not just the final answer.

Another approach is to integrate self-verification capabilities into the agents themselves. Instead of relying solely on an external verifier, agents could be trained to double-check their own work at critical junctures. This would reduce the burden on external systems and make the agents more robust in situations where no external verifier is available.

Finally, Garg highlights the importance of human-in-the-loop verification for high-stakes tasks. While automation is valuable, certain decisions may still require human judgment. A hybrid model—where AI verifiers flag issues and humans make the final call—could strike the right balance between efficiency and safety.

Industry Implications

The push for better verifiers has significant implications for the broader AI industry. As companies deploy AI agents in production environments, the reliability of these systems will be a key differentiator. Startups that can crack the verification problem may gain a competitive edge, while those that neglect it could face reputational damage from high-profile failures.

For the crypto and blockchain space, where AI agents are increasingly used for trading, compliance, and governance, the stakes are even higher. A single flawed decision in a long-horizon trading strategy could lead to substantial financial losses. This makes robust verification not just a technical nicety, but a business necessity.

Key Takeaways

Rayan Garg’s insights serve as a timely reminder that the AI industry&rsquos focus on model capability must be matched by an equal focus on verification. As agents take on longer and more complex tasks, the need for better verifiers will only intensify. Companies that invest in this area now will be better positioned to build trustworthy AI systems that users can rely on.

  • Long-horizon AI agents require verification systems that assess intermediate steps, not just final outputs.
  • Current verifiers are inadequate for dynamic, multi-step tasks.
  • Process-based verification, self-checking, and human oversight are promising solutions.
  • Reliable verification will be a key competitive advantage in the AI and crypto industries.

As the field evolves, one thing is clear: the future of AI depends not just on how smart the agents are, but on how well we can verify what they do.