Malleable Evals: Why Are We Evaluating Adaptive Systems with Static Tests? — Vincent Koc, OpenClaw

Vincent Koc, a contributor to OpenClaw and eval researcher at Comet who works with organizations from Uber and Netflix to UK banks, argues that the AI industry has a fundamental mismatch between how AI applications actually behave and how they are evaluated. The central thesis is what he terms 'eval calcification': evaluation frameworks built as static datasets are increasingly inadequate for agentic, adaptive AI systems that shift behavior as user bases, prompts, and toolchains evolve.

Koc traces the evolution of AI interaction paradigms from prompt engineering (2022-2023, described as 'bashing random words into an AI and hoping it improves'), through context engineering (RAG, tool calling, MCP-based agents), to what he calls 'intent engineering' — where models self-optimize based on a stated goal or end state rather than responding to rigid instructions. He points to Claude Code (OpenClaw harness), Codex, and similar systems as current examples of intent-based machines that actively adapt to the user. The ARC-I2 and ARC-I3 puzzle benchmarks are cited as evidence that large language models can do non-trivial pattern recognition that humans struggle with.

The talk defines the 80/20 breakdown at the core of the problem: 80% of what an agent does is stable and predictable, but 20% is fluid — it changes as customers change, as edge cases emerge, as the product evolves. That 20%, Koc argues, is precisely what destroys production systems and precisely what static benchmark suites are blind to. Conventional evaluation practice — hand-crafted question sets, offline regression suites, fixed compliance checks — was designed for static software and cannot track this drift.

The replacement Koc sketches has four components. First, intent-based outcome definitions: rather than asserting that a specific input must produce a specific output, evals should define the end state the agent is supposed to reach and evaluate whether it reached it — analogous to rubric-based grading rather than answer-key grading. Second, self-curating test suites: agents should mine their own traces to identify what has changed in user behavior and automatically update or extend their test cases, without human intervention. He references Andrej Karpathy's auto-research / auto-optimization pattern as a minimal Python proof-of-concept: set a goal, set a reward signal, let the system tune itself. Third, always-on online evaluation: rather than periodic offline batch runs, evaluation becomes a continuous service running against live traces. Fourth, telemetry-in-the-loop self-correction: harnesses that are aware of their own error telemetry, cost signals, and failure conditions can self-correct mid-run rather than waiting for a human to diagnose and re-deploy.

Koc distinguishes this from the popular dismissal of evals as 'dead.' His position is the opposite: in an era of intentful, self-optimizing agents, the need for evaluation is greater than ever — but the evaluation system itself must be treated as a living agent, not a point-in-time dataset. The benchmark is not a spreadsheet; it is a self-optimizing system that has to grow alongside the application it measures. He notes that a full demonstration of Comet's implementation of these ideas was not yet ready at the time of the talk but would be available in coming weeks.