This paper argues that as large language models acquire the capacity for sustained, multi-step, tool-mediated execution, the locus of engineering value migrates from authoring deterministic code toward governing probabilistic, autonomous behavior — and that this is a paradigm shift rather than a tooling upgrade, giving rise to a distinct professional archetype it calls the Agentic Engineer. Its central claim is one of symbiosis rather than substitution: the agentic engineer is constructed upon, and remains accountable through, classical engineering discipline.
The shift is characterized along three axes. The unit of work moves from deterministic code implementing a feature to a supervised agent workflow, so that engineering effort goes into shaping the operating environment — what an agent can perceive, which tools it may access, what modifications it may make, and when human intervention is required — rather than prescribing every computational step. The correctness model moves from binary assertion, verifiable locally through tests and specifications, to statistical evaluation under uncertainty, where the question becomes whether a system is sufficiently reliable under realistic operating conditions and whether its failure modes are acceptable. The accountability model moves from authorship to outcome ownership: agents may generate code, review pull requests and propose modifications, but do not assume responsibility for outcomes, so governance mechanisms such as auditability, approval workflows and intervention policies become architectural requirements rather than operational afterthoughts.
The paper states its own epistemic status directly: it is a position paper whose goal is to synthesize a rapidly evolving body of literature into a coherent argument and a set of testable predictions, drawing quantitative findings from the studies it cites and framing forward-looking claims explicitly as hypotheses.
Key Points
- A fourteen-dimension comparison contrasts the software engineer with the agentic engineer across core paradigm, primary output, unit of work, knowledge domain, tooling, architecture, design pattern, lifecycle, testing, debugging, security, accountability, success metrics and mindset. The paper is explicit that this table is an analytical scaffold rather than an empirically validated taxonomy.
- The paper deliberately foregrounds contested evidence rather than only supportive results: a lab experiment reporting a 55.8% reduction in completion time, three field experiments across 4,867 developers reporting 26.1% more completed tasks with the largest gains for less experienced developers, and a randomized controlled trial of 16 experienced open-source developers on their own mature repositories reporting a 19% increase in completion time — even though those developers had forecast a 20–24% speedup.
- It draws the argument of the paper from that juxtaposition: gains concentrate on well-scoped tasks and lower-context cohorts while high-context expert work can incur net costs, which is why it holds that oversight and calibrated judgment rather than raw automation are the load-bearing competencies, and why it treats the perception–reality calibration gap as itself a managed risk.
- Compositional Reliability is formalized: if an agent must complete n dependent steps each succeeding independently with probability p, end-to-end success is p to the power n, so a per-step rate of 0.95 yields roughly 0.36 over twenty steps. The paper notes this is an optimistic upper bound, since real errors correlate and cascade.
- Behavioral Drift is formalized as two distinct shifts — data drift in the marginal input distribution and concept drift in the conditional — and compounded by provider-side model updates, so that an agent has no stable terminal state and requires permanent stewardship.
- The paper reports measured attack rates for indirect prompt injection, citing a tool-integrated benchmark that found ReAct-prompted GPT-4 agents successfully attacked in roughly 24% of cases, and argues that defenses differ in kind from classical perimeter security: strict tool-permission scoping, output guardrails, sandboxed execution and human-in-the-loop checkpoints for destructive actions.
- It treats evaluation as the central artifact of agentic practice rather than an afterthought, while documenting that the dominant grading mechanism is itself imperfect — LLM judges exhibit position, verbosity and self-preference biases, and judge choice can reorder model rankings.
- Six predictions are advanced, each stated as a falsifiable hypothesis with the observation that would refute it: convergence of the two roles into a hybrid archetype; agents performing first-pass work across the lifecycle; effects remaining strongly heterogeneous; evaluation and reliability consolidating into a named discipline; the falling cost of modification turning one-way architectural decisions into two-way doors; and standards fluency becoming a hiring filter.
Notes
The paper maps agentic-engineering practice onto the SFIA 9 responsibility ladder, running from executing deterministic commands at Level 1 through constructing multi-agent state loops and evaluation pipelines at mid-levels to leading ISO/IEC 42001 and NIST AI RMF audits and governing model portfolios at the top. Its governance section draws on three complementary frameworks — ISO/IEC 42001 as an auditable AI management system, the IEEE 7000 series as a value-engineering process tracing stakeholder values to technical requirements, and the NIST AI RMF as a continuous govern–map–measure–manage life cycle — with the throughline that governance is integrated into architecture rather than appended to it.
Its open problems are evaluation validity, since LLM-as-judge introduces circularity and documented bias while trajectory metrics may reward spurious tool sequences; compositional reliability, since principled methods for bounding end-to-end reliability under correlated failures are lacking; security under agency; accountability attribution, which it describes as legally and ethically unsettled when an agent proposes and a human approves, especially across vendor model updates that silently alter behavior; workforce formation, where curricula and certification taxonomies lag practice; and sustained stewardship, where the cost model of permanent monitoring and re-alignment is not well understood at portfolio level.
The paper states its own threats to validity as a synthesis. The controlled-evidence base is small, recent and tied to specific tool generations, so the effect sizes should be read as time-stamped snapshots rather than stable constants; the fourteen-dimension comparison is an analytical construct; the forecasts concern a fast-moving field and are offered as falsifiable hypotheses. It says it has sought to mitigate selection bias by reporting disconfirming evidence, while noting a residual bias toward published, English-language literature.