This preprint, written by an author at Northeastern University and marked as under review, is a synthesis rather than an original study. It contrasts the in-editor completion tools that followed the 2021 release of GitHub Copilot — where, in the paper’s framing, the human remained the engineer and the model was an autocomplete with judgment — with agentic systems such as Claude Code, OpenAI’s Codex CLI, Google Jules, Devin, OpenHands and SWE-Agent, which read a repository, plan changes across files, run commands and tests, observe failures and deliver a committed change. Drawing on published material from industrial labs and academic groups, it offers four contributions: a synthesis of the main lines of work, a six-layer reference architecture, a contrast between the traditional software development lifecycle and an Agentic SDLC, and a consolidation of empirical evidence together with an agenda of open problems.
The paper’s central argument is that code generation is the wrong frame for the transition. The right frame, it proposes, is delegated execution under human supervision, and it expects the long-run winners to be those who invest earliest in the process, governance and skills that delegation requires. It describes its overall reading as cautiously optimistic.
Key Points
- The proposed reference architecture has six layers, with dependency running upward and feedback flowing downward: L0 the foundation model; L1 reasoning, memory and self-reflection; L2 the Agent-Computer Interface; L3 tools and environment; L4 orchestration, where single-agent loops and role-specialized multi-agent systems coexist; and L5 governance and safety. The author argues that L5 is currently the least mature layer and is becoming the bottleneck on enterprise deployment.
- Following industry frameworks, the paper distinguishes an Agentic SDLC, in which an orchestrator coordinates specialized sub-agents while a human supervises at intent, review and approval gates, from an SDLC merely accelerated by AI assistants. It emphasizes three structural differences: the unit of work shrinks to tasks an agent completes in minutes to hours; the developer’s role shifts from producing to orchestrating, reviewing and directing; and behavioral metrics such as agent acceptance rate, escalation quality and supervision burden displace, without entirely replacing, cycle time and defect rate.
- It argues that agentic systems break the classical SDLC’s assumption that behavior is fully specified at build time, because both the systems and the development process become stochastic: the same task may take a different path on different runs.
- Its compiled SWE-Bench Verified timeline, drawn from vendor reports and papers, rises from 1.96% in October 2023 to 78.4% in April 2026, while non-agentic retrieval-based systems plateau near 20%; the author reads this as the gain being dominated by scaffolding rather than raw model capability.
- On productivity it groups evidence into controlled experiments, which report the largest effects, field experiments and longitudinal team studies, and cautions that gains are unevenly distributed and that AI-assisted code may increase long-term technical debt.
- It observes convergence across the major programs: all expose a CLI- or IDE-resident agent with shell, file and test access, human approval at high-impact actions, some form of long-term memory, parallel or multi-agent execution for non-trivial tasks, and a story about safety and audit.
- The five open problems it names are evaluation beyond SWE-bench (long-horizon delegation, collaboration with reviewers, faithfulness to intent, absence of reward hacking); governance, safety and audit, including making human–agent responsibility mapping verifiable by auditors; the technical-debt hypothesis, since agents are biased toward producing more code and local fixes; skill redistribution, with experienced engineers capturing compounding gains while newcomers underperform; and the economics of attention, since if agents produce many plausible patches per hour, human review becomes the rate-limiting resource.
Notes
The paper’s quantitative claims are compilations of other work — vendor announcements, model cards, Anthropic Economic Index reports, and productivity studies on GitHub Copilot — rather than new measurements, and several of its figures (including its table of approximate SWE-bench Verified scores per program) are marked by the author as approximate. It also places major systems on a capability–autonomy plane and contrasts the strategic posture of the major industrial programs. The version extracted here is arXiv:2604.26275v1.