What makes agent-created code artifacts so hard to manage?

Among the three elements of agentic code — model capability, harness infrastructure, and agent-initiated artifacts — the survey flags the third as the one that "remains relatively underexplored." Agent-initiated code artifacts are the interactive objects an agent creates, executes, observes, revises, persists, and shares during a task: patches and tests authored over a live repository, interface commands synthesized against DOM trees, hypothesis-testing pipelines composed on the fly, executable policies and skill libraries revised in response to environment feedback. These appear across coding assistance, GUI/OS automation, scientific discovery, and embodied control — yet they sit outside the well-mapped territory of predefined infrastructure.

The open questions cluster around persistence and sharing. When an agent writes code that outlives the current step, what should persist and what should be discarded? When multiple agents share artifacts, how is consistent state maintained, and how is a useful artifact promoted from one-off scratch work to durable, reviewable infrastructure? The survey's listed open challenges — evaluation beyond final task success, verification under incomplete feedback, regression-free harness improvement, consistent shared state across agents, human oversight for safety-critical actions — converge on exactly this layer. The counterpoint is that some agent-authored code is genuinely disposable and over-engineering its lifecycle wastes effort. But this matters because the artifacts an agent creates may be where the next gains in autonomy and coordination live, and they are precisely what current harness engineering least understands.

— "Code as Agent Harness: Toward Executable, Verifiable, and Stateful Agent Systems", https://arxiv.org/abs/2605.18747