Code agent evaluation has been dominated for years by a simple metric: the final result. An agent is good if it passes a test suite and bad if it does not. That binary view, useful for comparing models quickly, leaves out the real experience of someone working with these tools in a development environment. Each run is a process made up of decisions, attempts, mistakes and corrections, and that whole journey determines whether an agent is genuinely productive or not. AgentLens addresses exactly this gap, proposing a trajectory-based evaluation instead of a result-based evaluation.
The proposal combines formal verification, trajectory reviews generated by language models and side-by-side comparisons. Formal verification provides an objective criterion when a clear check exists, such as the result of automated tests or the validity of a code transformation. Trajectory reviews, for their part, interpret the agent behavior during execution: how it follows instructions, how it selects and uses its tools, how it verifies its own work, how it recovers from mistakes and how it communicates with the person supervising it. Side-by-side comparisons make it possible to place two runs face to face and reason about which one is safer, more efficient or more understandable. The result is not an opaque score, but a readable explanation of why a specific score was obtained.
For a software development company like Q2BSTUDIO, trajectory-based evaluation has very valuable practical implications. When we build custom software that includes AI agents, we need to know not only whether the agent finishes the task, but whether it does so in a way that can be audited, maintained and explained to the client. An agent that passes a test but has overwritten data without warning, or has worked on the wrong branch, would be a risk in a real project. Trajectory evaluation makes it possible to detect these behaviors before they reach production.
It also fits with our vision of artificial intelligence applied to business processes. It is not about replacing people, but about creating assistants that know when to act and when to ask. For those assistants to be reliable, the measurement must include the quality of interaction, the clarity of explanations and the ability to correct course. AgentLens makes it possible to monitor all these aspects in a nightly pipeline, so that a new version of the agent can be compared with the previous one before deployment. It is a diagnostic and regression control tool, not just a ranking tool.
In terms of infrastructure, an evaluation system of this kind requires a certain technical maturity. Agent runs can consume significant resources, so it is advisable to have cloud AWS/Azure environments that allow launching tests in parallel and isolating each experiment. At Q2BSTUDIO we regularly work with these platforms to build continuous integration environments, and agent evaluation follows the same pattern: each run must be reproducible, traceable and destroyable. The cloud also simplifies credential management, a critical aspect because an agent with access to a repository must have least-privilege permissions and never expose secrets.
Cybersecurity also plays a relevant role. A poorly evaluated code agent can become a gateway for prompt injection attacks or sensitive information leakage. That is why, in our cybersecurity projects, we apply an additional layer of control to any AI-based automation. An agent trajectory is an audit log: it makes it possible to reconstruct what it did, with what data and with what consequences. Evaluating that trajectory is, in the end, a security measure.
On the other hand, AgentLens results can be integrated into Business Intelligence dashboards. Organizations that invest in code agents want to see trends: how the success rate evolves, what types of tasks generate more errors, whether trajectory reviews show poor communication patterns. With BI/Power BI tools it is possible to turn this data into dashboards that help decide where to focus improvement. At Q2BSTUDIO we help our clients design that observability layer, connecting agent evaluation with business metrics.
The philosophy of AgentLens is not exclusive to academia. As an open-source benchmark, it allows any team to adopt it, adapt it and run it in their own pipeline. That makes it especially attractive for companies that want to make evidence-based decisions rather than impression-based ones. Instead of only asking whether the test passed, the team can ask whether the agent worked as expected. The difference seems subtle, but in practice it changes the conversation completely.
The emphasis on trajectory also has a direct consequence on trust. Developers who use AI assistants need to understand what the agent has done before accepting its result. A binary score does not build trust; a detailed explanation does. When an agent can show that it ran the tests, fixed the issues found and communicated its decisions clearly, its work becomes much easier to review. That reviewability is the foundation for integrating agents into serious workflows.
In short, AgentLens represents a necessary step in the maturity of code agents. As more companies rely on AI assistants to write, review and deploy software, we need better instruments to understand their behavior. Trajectory evaluation, with its readable explanations and contextual comparisons, is probably the most sensible path. At Q2BSTUDIO we closely follow these proposals to apply them in custom software, cloud, cybersecurity and business intelligence projects, because we know that a useful agent is not only the one that gets it right, but the one that gets it right for the right reasons.
