Spec-Driven Development for Agentic Software Engineering: Harnessing Human-Agent Teamwork

A conceptual foundation for governing autonomous agents at team scale through specification-centric contracts and socio-technical harnesses

Jessica Diaz; Joaquin Gayoso; Andrea Cimminio; Jorge Perez· 2026· DOI 10.48550/arXiv.2609.00252

The core problem

Software engineering is transitioning from AI-assisted practices—exemplified by *vibe coding*, where assistants accelerate individual developers—toward **Agentic Software Engineering (ASE)**, in which autonomous agents are delegated goal-level tasks. Despite rising individual productivity, industry reports a **productivity paradox**: team throughput, review capacity, and stability degrade because team-scale software engineering discipline is neglected. The authors aim to establish the conceptual and methodological foundations of **Spec-Driven Development (SDD)** as an enabling discipline for ASE at team scale, and to characterize the *harness*—the technical and methodological mechanisms through which teams govern agent behavior. The central research question is: *How can specifications serve as the contract substrate between humans and agents to preserve accountability, verifiability, and transferability in agentic teams?*

Innovation

The paper presents three main results: **(i) Socio-technical model of SDD.** Specifications are elevated from documentation artifacts to *executable contracts* that bind human intent to agent behavior. This model reconstitutes, in specification-centric form, the contracts that vibe coding dissolves: **accountability**, **verifiability**, and **transferability**. **(ii) Operational characterization of the harness.** The harness is split into two layers: - **Technical harness** around the agent: sandboxing, tool access control, output validation, and runtime monitoring. - **Methodological harness** around the team: review gates, specification versioning, traceability matrices, and role definitions. Worked examples illustrate how these layers interact in practice. **(iii) Typology of five human-agent interaction patterns.** These patterns redefine the human role from coder to specifier, reviewer, orchestrator, auditor, and exception handler. The patterns are not mutually exclusive; teams may combine them depending on task criticality and agent autonomy level. The authors also note that the productivity paradox persists when only the technical harness is implemented without the me

Software engineering is transitioning from AI-assisted practices—exemplified by *vibe coding*, where assistants accelerate individual developers—toward **Agentic Software Engineering (ASE)**, in which autonomous agents are delegated goal-level tasks. Despite rising individual productivity, industry reports a **productivity paradox**: team throughput, review capacity, and stability degrade because team-scale software engineering discipline is neglected. The authors aim to establish the conceptual and methodological foundations of **Spec-Driven Development (SDD)** as an enabling discipline for ASE at team scale, and to characterize the *harness*—the technical and methodological mechanisms through which teams govern agent behavior. The central research question is: *How can specifications serve as the contract substrate between humans and agents to preserve accountability, verifiability, and transferability in agentic teams?*

The study conducts a **conceptual analysis** drawing predominantly on gray literature—ASE vision and roadmap papers, practitioner reports, talks, and tooling—because peer-reviewed evidence and a shared academic-industrial vocabulary are not yet established. The authors use a **comparative characterization** of the paradigm progression (from manual coding to AI-assisted to agentic) as a conceptual framing. They then derive:

Why it matters

The discussion centers on the **contractual nature of specifications**. In traditional software engineering, contracts (e.g., interfaces, tests, SLAs) ensure accountability and verifiability. Vibe coding erodes these contracts by prioritizing speed over explicit agreement. SDD restores them by making specifications the primary artifact of negotiation between humans and agents. The **dual harness** is presented as a socio-technical control system. The technical harness alone cannot prevent team-scale degradation because it does not address coordination, review capacity, or knowledge transfer. The methodological harness provides the organizational scaffolding—such as specification review boards, agent performance audits, and rollback procedures—that keeps agentic teams stable. The **five interaction patterns** imply a shift in engineering education and team composition. Humans must learn to write precise, testable specifications and to audit agent behavior. The authors caution that the evidence base is immature; the typology and harness model are hypotheses to be empirically validated. They outline a research agenda including controlled experiments, longitudinal case studies, and the development of shared metrics for agentic team performance. A key implication is that **SDD is not a tool but a discipline**. It requires cultural change: from celebrating individual speed to valuing team-level throughput, review quality, and long-term maintainability. The paper concludes that without SDD, ASE will amplify the productivity paradox rather than resolve it.

Who should read this

CS practitioners and researchers

Opening member content…