The four-stage evolution of agentic engineering is a roadmap proposed in Agentic Software: How AI Agents Are Restructuring the Software Paradigm, based on what the paper describes as current capabilities and trajectories. Each stage is characterised along four axes: what the agent can do, the key technologies that enable it, what the human role becomes, and which systems are representative of it.
Usage
Stage I, Tool-Augmented (2023–2025), is the mode the paper identifies as currently dominant. Agents serve as assistants within human-led workflows, handling code completion, single-issue fixes and simple script generation through in-context learning and retrieval. The human is author and reviewer. GitHub Copilot and Claude Code are named as representative. The limitation is that the human must still decompose problems, design architecture and verify correctness.
Stage II, Single-Task Autonomous (2025–2027), has agents own complete tasks from specification to deployment — end-to-end feature building, debugging, basic system maintenance — through planning, tool use and self-correction. The human shifts to intent architect and auditor: from doing, to specifying what to do and verifying what was done. Devin and OpenHands are named as demonstrating that agents can autonomously navigate codebases, implement features and submit pull requests.
Stage III, Multi-Agent Teams (2026–2029), has specialised agents coordinate as teams mirroring human engineering organisations: a product-manager agent translating business requirements into technical specifications, architect agents designing structure, developer agents implementing components, QA agents testing and validating. The enabling technologies are shared memory, role specialisation and orchestration; shared memory and observability become critical infrastructure. The human role becomes PM plus architect plus auditor, and LangChain orchestration and MetaGPT are named as representative.
Stage IV, Self-Evolving Ecosystems (2028+), has agents improve their own architectures, spawn specialised sub-agents for new problem domains, and adapt to environmental change without human intervention. At this stage, the paper argues, the distinction between software and agent dissolves entirely — the agent is the system and evolves continuously — and human involvement shifts to meta-level governance: setting ethical boundaries, defining value functions and ensuring alignment. Its representative systems are given as prospective rather than existing.
When It Applies
The scheme is a forecast, and its stage boundaries are the paper’s own reading of a trajectory rather than thresholds with a stated measurement behind them. The date ranges given for the stages overlap, which fits a roadmap describing what becomes possible rather than what replaces what; the paper does not comment on the overlap.
It assumes the capability trend it extrapolates from continues. The paper’s own empirical section is the place to read against that assumption: it reports the EvoClaw benchmark’s finding that success rates collapse from above 80% on isolated tasks to at most 38% on continuous software evolution, and treats the gap between those two numbers as quantifying the distance still to be covered. On the paper’s own account that gap is not fundamental — it reflects limitations in context management, memory architecture and verification that are active research areas — but the paper states directly that fully autonomous software development will require several more years of concentrated research before it is reliable in production.
Used as a positioning device the scheme is one of several. It is a distinct scheme from Agentic Autonomy Levels and SE Autonomy Levels (SE1.0–SE5.0), which cut the same territory differently; nothing in this source reconciles it with either, and a reader comparing them should treat the stage numbers as internal to this paper.
Related Terms
- Agentic Software: How AI Agents Are Restructuring the Software Paradigm — the paper that proposes this roadmap
- Agentic Engineering — the discipline whose evolution the stages describe
- Agentic Autonomy Levels — a separate scheme over similar ground
- SE Autonomy Levels (SE1.0–SE5.0) — another such scheme
- EvoClaw — the benchmark whose results the same paper uses to locate current capability