Calibrated-Trust Orchestration: A Meta-Layer Framework for Coordinating Task-Level AI Agents Across the DevSecOps Continuum

Abstract

Generative and agentic artificial intelligence (AI) tools are being adopted throughout the DevSecOps continuum, yet current deployments remain fragmented: each tool operates on an isolated task, exposes its own ad-hoc confidence signal, and is trusted - or distrusted - by developers on the basis of anecdote rather than measurement. Recent industry synthesis has identified this fragmentation, together with the absence of a mechanism for recalibrating human trust in AI outputs over time, as a primary barrier to safe adoption of AI-assisted software engineering. This paper proposes Calibrated-Trust Orchestration (CTO), a meta-layer framework that sits above individual tasklevel AI tools and agents and governs the autonomy each is granted using a continuously updated, task-conditioned trust score. The score combines historical acceptance rate, humanoverride frequency, and a hallucination-risk proxy specific to each pipeline stage (plan, code, build, test, release, deploy, operate, monitor). Rather than treating every AI suggestion identically, CTO assigns each agent to one of three autonomy tiers - autonomous, suggest-with-rationale, or human-gated-and migrates agents between tiers as evidence accumulates. We present the formal trust-score model, the tier-assignment algorithm, an architecture for embedding CTO into an existing CI/CD toolchain, and a worked illustrative scenario showing how the framework would behave across a release cycle. The proposal directly targets two gaps identified in prior work: the lack of cross-tool coordination and the absence of a structured human/machine recalibration loop. We close with a discussion of validation requirements and threats to the approach.