(Over)Reliance on Test Agents in AI-Assisted Software Testing¶

🕒 Published (v1): 2026-07-20 13:23 UTC · Source: Arxiv · link

Why this paper was selected

Empirical study of over-reliance on test agents; calibration signal for agent harness quality gates

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This position paper argues that AI-based test agents introduce two compounding risks in software testing: an agency problem (engineers cede cognitive control over test design) and an assurance problem (agent-produced artifacts are accepted as evidence without scrutiny). The authors propose a conceptual framework for studying overreliance and enumerate 12 domain-specific overreliance modes.

Problem¶

As test agents become capable of generating, selecting, executing, delegating, and regenerating tests autonomously, human engineers shift from direct test design to supervisory roles. Existing literature on automation bias and ethical AI testing flags over-trust as a risk, but no taxonomy of overreliance modes specific to agentic software testing exists, nor is there a methodological framework for empirically studying how and where overreliance occurs in such workflows.

Method¶

The paper develops its argument through three theoretical lenses applied to agentic test workflows: 1. Software testing as cognitive problem-solving — testers perform goal formulation, technique selection, adequacy judgment, and result interpretation; agent outputs can silently replace each cognitive step. 2. Test agents as adaptively autonomous entities — agents maintain states (Idle, Interact, Execute, Regenerate, Out of Order), delegate subtasks, and interact peer-to-peer; their actions unfold over time rather than as static artifacts. 3. Test design argumentation — tests can be structured as explicit claim–reason–evidence–warrant chains, making agent reasoning inspectable and challengeable.

From these lenses, the paper derives a conceptual framework (Figure 2) specifying an iterative data-collection loop: model agent behavior → operationalize constructs → deploy adaptive/argumentative test agents → measure reliance outcomes → refine the taxonomy. Reliance measures include acceptance rate of incorrect claims, frequency of challenged outputs, evidence-inspection behavior, and human-override counts.

Key Contributions¶

Framing overreliance on test agents as simultaneously a supervisory-control problem and an assurance problem.
A taxonomy of 12 overreliance modes (Table 1): Test Goal, Test Strategy, Claim, Reason, Evidence, Test Argument, Test Execution, Oracle, Delegation, Regeneration, Escalation, and Maintenance overreliance.
A conceptual framework for future empirical studies combining open-source and industrial settings.
Definition of a test agent as a formal construct distinct from generic AI tooling: an autonomous/semi-autonomous agent that observes SUT changes, maintains testing goals, generates/selects/executes artifacts, interacts with other agents or humans, and produces assurance evidence.
An illustrative industrial-control-system scenario demonstrating how multiple overreliance modes co-occur in a multi-agent delegation chain.

Results¶

No empirical results are reported; this is a position/framework paper. All claims are conceptual or interpretive. The authors explicitly note: "Without case study evidence, claims about the overreliance remain interpretive."

Limitations¶

No empirical validation; the overreliance taxonomy is the author's interpretive construction.
The framework describes a data-collection plan but does not instantiate or pilot it.
Only one author perspective is represented; the proposed modes offer "only a partial view."
The adaptive/argumentative test agent architecture needed to operationalize the framework does not yet exist in a deployable form.
The industrial scenario is illustrative, not drawn from observed data.

Relevance to Harnesses / Meta-Harnesses¶

This paper directly engages the meta-harness concern: in a multi-agent testing workflow, the harness itself (the layer that orchestrates agent state, delegation, evidence aggregation, and regeneration) is the locus where overreliance silently propagates. The 12-mode taxonomy maps onto canonical harness responsibilities—goal dispatch, subtask delegation, result summarization, oracle evaluation, and suite maintenance—making it a diagnostic checklist for anyone designing or auditing an agentic test harness. The "semi-executable stack" concept (combining executable code, prompts, agent workflows, evaluation harnesses, policies, and human judgment) aligns precisely with the meta-harness abstraction in the broader literature, and the proposed empirical framework offers a methodology for measuring where human oversight erodes in such stacks.