From Helpful to Trustworthy: LLM Agents for Pair Programming

Doctoral research proposes a systematic framework for multi-agent LLM pair programming that improves code reliability and auditability through externalized intent and iterative validation. The study addresses critical gaps in how AI coding agents can produce trustworthy outputs aligned with developer objectives across testing, implementation, and maintenance workflows.

Current LLM-based coding agents face a fundamental credibility problem: they generate plausible code that may diverge from actual developer intent and lack transparent evidence trails for review. This research tackles a real pain point in AI-assisted development—the gap between helpful outputs and trustworthy, production-ready code.

The doctoral work builds on growing recognition that LLM agents need structured workflows rather than isolated prompts. By proposing multi-agent systems that externalize intent through formal specifications and use automated tools for iterative validation, the research addresses how to move beyond trial-and-error prompting toward systematized development practices. The three proposed studies—translating requirements to formal specs, refining implementations with automated feedback, and managing maintenance tasks—reflect practical development challenges where audit trails and reproducibility matter.

For the developer ecosystem, this framework could shift how teams integrate AI into engineering workflows. Rather than using LLMs as code generators alone, the approach treats them as collaborative agents within verifiable development pipelines. Tools like solver-backed counterexamples and preserved behavioral validation create accountability mechanisms absent from current LLM pair-programming implementations.

The broader impact extends to AI infrastructure vendors and enterprise adoption. As organizations demand trustworthy AI coding assistants, research establishing when multi-agent workflows increase trust becomes commercially relevant. Success here could accelerate enterprise adoption of AI development tools, but only if the research produces actionable patterns that practitioners can implement without substantial infrastructure overhauls.

• →Multi-agent LLM workflows with externalized intent and automated validation can improve code trustworthiness beyond single-agent approaches.
• →Formal specifications and automated feedback mechanisms create auditable evidence trails critical for production code reliability.
• →Developer intent misalignment remains a core challenge requiring systematic solutions beyond prompt engineering.
• →Structured validation across testing, implementation, and maintenance preserves behavioral correctness as codebases evolve.
• →Enterprise adoption of AI coding assistants depends on frameworks that ensure auditability and maintainability, not just code generation speed.