AI-native software engineering is the term From Determinism to Delegation: AI-Native Software Engineering and the Evolution of the Agentic Engineer gives to the paradigm it argues is emerging as large language models acquire the capacity for sustained, multi-step, tool-mediated execution. Its claim is that this is not an incremental tooling upgrade applied within existing practice but a change in how software work is structured, how system quality is evaluated, and how responsibility is assigned. The paper argues the three changes are connected — a change in one tends to influence the others — and together describe a transition from deterministic implementation toward engineering centred on delegation, supervision and probabilistic behaviour.
Usage
The paradigm is characterised along three axes.
The unit of work. Classically the primary unit is deterministic code implementing a feature, service, API or platform component, and the process emphasises decomposing systems into modules with specified behaviour and verifying each artifact. In this paradigm the unit becomes the supervised agent workflow. The design questions shift to what an agent can perceive, which tools it may access, how it reasons about tasks, what modifications it is allowed to make, how outputs are evaluated, and under which conditions human intervention becomes necessary. Rather than prescribing every computational step, the engineer shapes the operating environment in which the system acts — so engineering effort moves from constructing individual behaviours toward constructing the conditions under which desirable behaviours are likely to arise.
The correctness model. Classical systems rely on binary correctness verifiable locally through assertions, tests and formal specifications. Agentic systems often operate where many tasks admit no single deterministic output, so correctness becomes statistical and system-level, measured through evaluation pipelines reporting task success rate, faithfulness, tool-use accuracy and hallucination frequency. The paper’s illustration is that a 94% task success rate may be entirely acceptable in one setting and unusable in another, so the central question shifts from “is the system correct?” to “is the system sufficiently reliable under realistic operating conditions, and are its failure modes acceptable?”.
The accountability model. In specification-driven development accountability follows from authorship. Here, agents may generate code, review pull requests, summarise incidents, identify defects or propose modifications, but do not assume responsibility for outcomes; human engineers retain ownership and remain accountable for quality, security and production readiness. The paper’s formulation is that responsibility does not disappear as autonomy increases — it becomes concentrated around oversight and outcome ownership, which is why governance mechanisms such as auditability, approval workflows and intervention policies become architectural requirements rather than operational afterthoughts.
The proposition the paper compresses all this into is that software engineering builds the system while agentic engineering builds the agentic system that helps build, operate and evolve the system.
When It Applies
The paradigm as described assumes agents capable of sustained multi-step execution against real tools, and a setting where outcomes can be evaluated statistically rather than asserted. It does not claim to displace classical practice: the paper is explicit that autonomous systems still depend on deterministic foundations — reliable interfaces, secure infrastructure, high-quality data, verifiable testing environments — and frames the relationship as complementary and mutually dependent rather than successive.
Its own evidence is where the paper sets the limits, and it foregrounds the disagreement rather than resolving it. Against a lab experiment reporting 55.8% faster completion on a bounded task, and three field experiments across 4,867 developers reporting 26.1% more completed tasks with the largest gains for less experienced developers, it sets a randomized controlled trial of 16 experienced open-source developers working on their own mature repositories, which found early-2025 AI tools increased completion time by 19% — while those same developers had forecast a 20–24% speedup. The reading the paper draws is that gains concentrate on well-scoped tasks and lower-context cohorts while high-context expert work can incur net costs, which is the basis for its thesis that oversight and calibrated judgment rather than raw automation are the load-bearing competencies. It treats the gap between perceived and actual benefit as itself a risk to be managed.
Two further limits are structural rather than empirical. Compositional Reliability means multi-step autonomy degrades multiplicatively, so high per-step accuracy does not carry over long horizons. Behavioral Drift means there is no terminal state: an agent requires permanent stewardship, so “done” is obsolete as a concept and the cost model of continuous monitoring and re-alignment is, the paper says, not yet well understood at portfolio level.
As a term it rests on a single position paper that presents itself as such — a synthesis of a rapidly evolving literature into an argument and a set of falsifiable predictions, drawing its quantitative findings from the studies it cites rather than from measurements of its own. The paper notes its controlled-evidence base is small, recent and tied to specific tool generations, and asks that the effect sizes be read as time-stamped snapshots rather than stable constants.
Related Terms
- From Determinism to Delegation: AI-Native Software Engineering and the Evolution of the Agentic Engineer — the paper that proposes this term
- Agentic Engineer — the professional archetype this paradigm is argued to produce
- Supervised Agency Spectrum — how the oversight this paradigm centres on is graduated
- Compositional Reliability — one of the two structural limits above
- Behavioral Drift — the other
- Agentic Engineering — an adjacent and separately-sourced term for the discipline