Skip to content

From Determinism to Delegation: AI-Native Software Engineering and the Evolution of the Agentic Engineer

🕒 Published (v1): 2026-06-27 07:43 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

A position paper arguing that LLM-powered autonomous agents constitute a full paradigm shift in software engineering—not a tooling upgrade—giving rise to the "Agentic Engineer" whose primary artifact is the supervised agent workflow, not deterministic code. The shift is characterized along three axes: unit of work (function → agent workflow), correctness model (binary assertion → statistical evaluation), and accountability model (authorship → outcome ownership). Central thesis: disciplined oversight, not raw automation, is the load-bearing competency.

Problem

Traditional software engineering assumes deterministic execution and binary correctness verification, but agentic systems produce probabilistic, emergent behavior that cannot be verified by assertions, static analysis, or coverage metrics. No unified framework exists for characterizing the competency gap, risk surface, or accountability structure this introduces for practicing engineers.

Method

Literature synthesis of peer-reviewed work published largely after 2022, combined with three analytical constructs: (1) a fourteen-dimension comparison table between the Software Engineer (SWE) and Agentic Engineer (AE); (2) formalization of agent mechanisms—the reasoning–acting loop as a POMDP policy \(a_t \sim \pi(a \mid s_t)\), scaled dot-product attention cost, compositional reliability decay \(P_\text{success}(n) = p^n\), and distributional drift; (3) mapping of emergent competency requirements onto SFIA 9 levels and governance standards (ISO/IEC 42001, NIST AI RMF, IEEE 7000). No new experiments are conducted; quantitative claims are drawn from cited studies.

Key Contributions

  • Three-axis paradigm-shift framing: unit of work, correctness model, accountability model.
  • Fourteen-dimension SWE-vs-AE comparison table (paradigm, output, tooling, architecture, testing, security, accountability, mindset, etc.).
  • Formalization of compositional reliability decay: at \(p=0.95\) per step, \(P_\text{success}(20) \approx 0.36\), motivating trajectory-level evaluation and HITL checkpoints.
  • Characterization of behavioral drift as data drift (\(P_\text{prod}(x) \neq P_\text{train}(x)\)) and concept drift (\(P_\text{prod}(y|x) \neq P_\text{train}(y|x)\)), requiring permanent TEVV stewardship.
  • Model Context Protocol (MCP) framed as the technical seam replacing \(N \times M\) bespoke integrations with a single client–server contract.
  • SFIA 9 competency ladder mapped to agentic-engineering practice across seven responsibility levels.
  • Six falsifiable predictions and an open-problems agenda on evaluation, security, governance, and workforce formation.

Results

Synthesized controlled evidence (paper makes no original empirical claims): - Lab experiment (N=95): −55.8% task completion time with AI assistant [Peng et al. 2023]. - Three field RCTs (N=4,867, Microsoft/Accenture/Fortune 100): +26.1% tasks completed; largest gains for less experienced developers [Cui et al. 2026]. - Telemetry + survey (~2,000 developers): tool-acceptance rate tracks self-reported perceived gain [Ziegler et al. 2024]. - RCT of expert OSS developers (N=16) on their own mature repositories: +19% completion time (net slowdown), despite developers forecasting a 20–24% speedup [Becker et al. 2025]. - Compositional reliability: \(p=0.95\) per-step success decays to \(\approx 0.36\) at \(n=20\) steps under the independence assumption (upper bound; correlated errors make it worse in practice).

Limitations

  • Position paper with no original empirical study; all quantitative findings are cited from others.
  • The fourteen-dimension comparison is explicitly an "analytical scaffold rather than an empirically validated taxonomy."
  • Controlled evidence on productivity is heterogeneous and contested; the key expert-developer RCT has N=16, limiting generalizability.
  • The claim "senior engineers adapt more readily" is acknowledged as hypothesis, not settled finding.
  • Forward-looking predictions are unfalsified by construction; the paper cannot distinguish which fraction of the paradigm shift is already realized versus speculative.
  • Compositional reliability formula assumes step independence—acknowledged as optimistic—but no tighter bound is derived.

Relevance to Harnesses / Meta-Harnesses

This paper provides a conceptual architecture for what harnesses and meta-harnesses are: the MCP client–server contract, orchestration topology (router–worker, hierarchical multi-agent, plan-and-execute), HITL gate placement, context engineering pipeline, and TEVV evaluation loop are precisely the components a harness designer must implement. The compositional reliability analysis (\(P_\text{success}(n) = p^n\)) formalizes why harnesses need trajectory-level checkpointing rather than per-step accuracy, directly motivating recovery and escalation hooks. The paper's treatment of behavioral drift under provider model updates identifies a meta-harness concern: harnesses must implement continuous statistical monitoring, not one-time integration testing, because the underlying model has no stable terminal state.