Dark Factory: How OpenClaw Ships Faster Than You Can Read the Diff — Vincent Koc
Vincent Koc argues that static benchmarks are failing in the era of adaptive AI. He proposes a shift from static testing to 'malleable evals,' where agents self-optimize and curate their own test suites based on user intent and production data, treating evaluation as a living, evolving system.
The Problem: Static Evals for Dynamic Systems
Traditional evaluation methods for AI are fundamentally broken because we treat dynamic, adaptive AI applications as if they are static software. While software engineering has long embraced concepts like chaos engineering and observability to understand system limits, the AI space remains hyper-fixated on static benchmarks and handcrafted datasets. This approach creates a significant gap, as AI systems, especially agentic ones, are not static. Their behavior drifts, their harnesses rewrite themselves, and they adapt to user interactions over time.
This creates an “eval calcification” problem, where our testing suites become increasingly rigid and irrelevant as the systems they are meant to measure evolve. The core issue is that a point-in-time snapshot cannot effectively measure a system that is constantly changing.
The Evolution of AI Engineering and its Impact on Evals
The challenge of evaluation has compounded with the progression of AI development methodologies:
-
Prompt Engineering: This early phase involved “bashing random words into an AI” to see what worked. It was an unstructured process of discovery, making systematic evaluation difficult.
-
Context Engineering: With the rise of RAG and tool-calling, systems became more structured. This allowed for more relevant evaluations, as individual components (like a specific tool) could be broken out and tested in isolation. This provided more steerability but didn’t solve the problem of evaluating the holistic, adaptive system.
-
Intent Engineering: This is the current frontier, where models are powerful enough and harnesses are sophisticated enough for agents to self-optimize based on user intent. Systems like Open Claw, Claude, and Codex adapt to provide personalized experiences. This makes evaluation exponentially more complex: how do you create a standardized test for a system that behaves differently for every user? This complexity, however, increases the need for robust evaluation and observability, rather than making them obsolete.
A New Paradigm: Malleable and Intent-Based Evals
The solution is to shift the mindset from static testing to a living, adaptive evaluation framework. Instead of starting with a fixed dataset, we should define the desired end state or intent-based outcome and allow the system to optimize towards it.
This new paradigm is built on several key principles:
- Define Abstract Outcomes: Move beyond simple question-answer pairs. Evals should measure more ambiguous, intent-focused qualities like an agent’s “personality” or its ability to handle “ambiguity,” potentially using rubrics similar to how art is evaluated.
- Self-Curating Test Suites: Empower agents to curate their own evaluation suites from production traces. If an agent detects that its customer base has changed and is asking new types of questions, it should automatically generate new tests to reflect this behavioral drift.
- Always-On, Online Optimization: Evaluation should not be a pre-deployment step but a continuous, “always-on” service that monitors the system in real-time.
- Telemetry-in-the-Loop: Agentic harnesses should be aware of their own telemetry—errors, costs, latency—and use this data to self-correct. If a particular pathway is error-prone or expensive, the system can learn to avoid it.
Managing the Critical 20%
In any production system, roughly 80% of interactions are predictable and stable. However, the 20% of edge cases are what cause the most significant business problems and user frustration. The goal of a malleable evaluation system is to create an agentic process that specifically monitors, manages, and adapts to this critical 20%.
Ultimately, we must stop treating evals as static datasets. Instead, they should be viewed as living systems—like code, software, or an agent in their own right. An evaluation framework must be a self-optimizing, growing solution that evolves in lockstep with the AI it measures.