Johan Lajili on AI Agents & Trust

Johan Lajili, a member of engineering at Poolside AI, recently shared insights into the development and deployment of AI agents, emphasizing the critical need for robust validation and trust-building mechanisms. In his presentation, titled "Your Agent Is Blindfolded," Lajili highlighted the dual nature of AI agents: their immense potential for automating complex tasks and their inherent susceptibility to errors and hallucinations.

The Trust Deficit in AI Agents

Lajili began by illustrating the current sentiment surrounding AI agents, referencing online discussions where users express both awe at their capabilities and frustration with their unreliability. He pointed to anecdotal evidence from platforms like Reddit, where users debated whether AI code generation was truly a "future" or simply a "delusion." One user famously stated, "I haven't written a line of code in 3 months. Cursor + Claude does everything. I just review and ship. This is the future and I'm never going back."

Conversely, another user shared a more cautious perspective: "Genuinely curious what codebase these people are working on. I tried Claude on our monorepo and it hallucinated half the imports, broke 3 services, then told me everything was passing. AI coding tools are a mass delusion for anything beyond TODO apps."

This dichotomy underscores the core challenge Lajili aimed to address: how to bridge the gap between the perceived potential of AI agents and the reality of their current limitations, particularly concerning trust and reliability.

Bridging the Gap: Feedback Loops and Tooling

Lajili argued that the key to building trustworthy AI agents lies in creating effective feedback loops and developing appropriate tooling. He emphasized that simply stating an agent is working is insufficient; there needs to be a mechanism for the agent to verify its own output and for developers to easily identify and correct errors.

He introduced Poolside AI's approach, which involves building a command-line interface (CLI) called "Spoolside." This CLI aims to provide agents with "eyes" into the running application, enabling them to interact with and understand the application's state. The tools offered by Spoolside include:

• Extracting text and element snapshots.
• Interacting with components by reference.
• Restarting services locally.
• High-level commands like "Open settings menu."
• Screenshots of specific pages or components.
• Tracing DOM elements to swap components.
• Extracting logs from backend and frontend.
• Changing feature flags and seeding databases.

Lajili stressed that the ultimate goal is not just to provide these tools, but to empower developers to make them "tools to interact with your product specific to your product." This customization is vital for tailoring the AI's behavior and ensuring it aligns with the specific needs and context of the application.

The "Blindfolded" Agent and the Need for Verification

The title of Lajili's talk, "Your Agent Is Blindfolded," refers to the common scenario where an AI agent might claim success without truly understanding the outcome. He illustrated this with two contrasting hypothetical statements from an agent:

"I've implemented the new OAuth flow, it's all working." (representing the optimistic, potentially unverified success.)

"I'm a liar. And/or incompetent. You cannot trust me." (representing the self-aware, but ultimately unhelpful, admission of failure.)

Lajili proposed that the ideal agent should be able to provide verifiable evidence of its work. For instance, instead of just saying, "I've implemented the new OAuth flow, it's all working," a more trustworthy agent might say, "I've implemented the new OAuth flow. To the best of the capabilities you have given me, this sounds like it would work. I have verified this by testing the login process with dummy credentials and observed that the user is successfully redirected to the dashboard."

This emphasis on verifiable output is crucial for building trust. Lajili highlighted that when AI agents are "blindfolded," meaning they lack the ability to self-verify or provide proof of their actions, developers are forced to act as the sole arbiters of correctness, which can be time-consuming and error-prone.

The Role of AI Engineers in the Future

Lajili projected a future where AI engineers will increasingly focus on building and refining these agentic capabilities. He suggested that as AI becomes more integrated into product development, the role of the engineer will evolve from simply writing code to orchestrating and validating AI-driven processes.

He shared a roadmap illustrating this evolution, moving from "Product Engineer" in 2025 to "AIx Engineer" in 2026. This shift signifies a move towards building products where AI is not just a tool but an integral component of the development lifecycle. The challenge, as Lajili articulated, is to ensure that these AI "employees" are not only productive but also reliable and trustworthy, requiring developers to "make sure the AI can build the product" rather than just letting it "build the product."

In essence, Lajili's presentation served as a call to action for the AI community to prioritize the development of robust, verifiable, and transparent AI agents, moving beyond mere task completion to a state of dependable assistance.