LLM Agents in Software Development:
A Systematic Mapping Study of Capabilities, Failure Modes, and Governance Guardrails

Claims about the benefits of AI in the software development lifecycle are often based on simplified benchmarks that underrepresent collaborative
work and long-term maintenance.

We present a systematic mapping study synthesizing 1,094 primary studies to provide a phase-aligned analysis of where LLMs and multi-agent systems deliver stable value, where they fail, and why. Our mapping reveals an uneven research landscape concentrated in code generation
and testing, with governance and deployment receiving comparatively little attention. We identify recurring failure modes: illusion of correctness
in generated code, fragile test suites, overly permissive CI/CD configurations, and code origin and license compatibility risks.

To mitigate these, we propose technical and procedural guardrails grounded in NIST SSDF and SLSA, including strict role separation between generators and independent verifiers and policy-as-code gateways. We complement the mapping with a practitioner reflection identifying where LLM tools provide stable value and which organizational controls reduce operational risk.

DOI: 10.1109/INES69513.2026.11661226
‍