AI-Driven Autonomous Software Engineering: Leveraging Agentic Systems for End-to-End Software Development and Testing
Abstract
ABSTRACT: The advent of Large Language Models (LLMs) capable of source-code synthesis has catalyzed a paradigm shift from static code-completion assistants toward autonomous, goal-directed software engineering agents [1], [2]. Unlike single-shot code generators, agentic systems integrate planning, external tool invocation, persistent memory, and iterative self-correction to autonomously traverse the full software development lifecycle (SDLC) — from requirement decomposition to code synthesis, automated test generation, execution-driven debugging, and deployment [4]. This review paper presents a comprehensive survey of agentic architectures for autonomous software engineering, encompassing single-agent reasoning-acting loops, multi-agent role-specialized pipelines (planner, coder, tester, reviewer), and orchestration frameworks. We propose and empirically evaluate a Test-Feedback-Driven Multi-Agent (TFD-MA) pipeline that couples iterative code refinement with automated unit-test execution feedback. Benchmarked against SWE-bench-style issue-resolution tasks, the proposed pipeline achieves substantially higher resolution rates and lower defect-escape rates than single-agent baselines, at the cost of increased inference latency and token expenditure. We further identify critical research gaps in long-horizon context retention, verification robustness, and multi-agent coordination overhead, outlining strategic future directions for trustworthy autonomous software engineering.
KEYWORDS: Agentic AI, Autonomous Software Engineering, Large Language Models, Multi-Agent Systems, Automated Code Generation, Test-Driven Development, AI Pair Programming, Software Development Lifecycle Automation.
Full Text:
PDF 67-78Refbacks
- There are currently no refbacks.
Copyright 2017 ManTech Publications. All Rights Reserved