Agentic Software: How AI Agents Are Restructuring the Software Paradigm
Zhenfeng Cao
cs.SE, cs.AI
2026-06-04
Position paper: the agent is the software, with logic generated at runtime. No new experiments. Sharpest citation is EvoClaw: >80% isolated, at most 38% under continuous evolution.
Traditional software encodes decision logic in static source. When requirements move, a person has to find and change that logic. This position paper, from an investment firm in Shenzhen, claims LLM agents change the object: the model plans, writes code, calls tools, and discards the code, so what persists is the agent's capability. Chinese discussion has read the genre as an obituary for software engineering. The text itself says the discipline expands rather than ends.
The practice it rejects is "AI → software → result": models help humans write conventional systems faster, while design, integration, and release stay on the human critical path and the delivered artifact keeps the same complexity ceiling.
Two tuples. A traditional system is S=(C, D, E) with D a static set of deterministic rules. An agent is A=(model, tools, memory, planner) stepping at = M(st, memory), s{t+1} = exec(at). Generated code is a transient instrument of reasoning.
The complexity argument is an upper bound: n components can realize 2^{C(n,2)} interaction topologies, human cognitive capacity is treated as constant, so the traditional path becomes uneconomic once the search exceeds that capacity, while agents outsource traversal to models whose capacity grows with compute. The paper notes that real systems do not realize every topology. Delivery history is told as three generations: licensed on-prem (complexity on the user), SaaS (on the vendor), AaaS (understanding and decision-making on the agent, priced on outcomes). Agentic Engineering is framed as an expansion: the artifact is the agent system, and humans become intent architects, coordinators, and auditors. A four-stage roadmap runs from tool-augmented coding in 2023–2025 to self-evolving ecosystems after 2028.
These are definitions and a narrative. There is no new training or evaluation protocol.
Every number is a citation.
| Source | Figure | What it is |
| Lingma SWE-GPT 72B | 30.20% on SWE-bench Verified | Near GPT-4o's 31.80% at the time |
| Same 7B variant | 18.20% | 22.76% relative over Llama 3.1 405B |
| LangChain blog pilot | 93% faster root-cause identification | 20+ enterprise debugging workflows; 200 engineering hours claimed in a month |
| EvoClaw (12 frontier models, 4 frameworks) | >80% isolated, ≤38% continuous | Sustained evolution across commits, errors accumulate |
The EvoClaw cliff is unpacked as context drift, error propagation, no model of technical debt, and tests that miss semantic faults. Hermes Agent is presented as a production system with self-patching Skills and subagent delegation, and the paper claims more than 179,000 GitHub stars. All of this is second-hand.
The usable takeaway is not "software has been redefined." It is the gap between isolated issue fixing and continuous evolution, 80% down to 38% in the cited benchmark. Work with clear success criteria, a bounded scope, and existing tests is the part that is agent-ready; long-horizon repo maintenance is not. Shifting humans toward intent, evaluation, and governance matches how many teams already operate, and does not require buying the asymptotic story.
Read as a systems paper it disappoints: no new benchmark, no ablation, no training setup to reproduce. A vendor blog and a star count are not controlled comparisons.
"First-principles necessity" overclaims. 2^{C(n,2)} is an unrealized upper bound, by the paper's own caveat, and does not imply agents are the only exit. Modularization, type systems, and formal methods are written off as constant-factor relief without a proof that the asymptote is what binds real projects. A 30% SWE-bench number and an orchestration blog cannot carry a qualitative break. The 38% EvoClaw cap, which the later sections accept, sits in tension with the inevitability rhetoric of the opening.
The empirical section is a collage: Karpathy's Software 2.0, ReAct, chain-of-thought, agent surveys, MetaGPT. There is no conference venue on the preprint. Treat it as a position essay. Criticism of "the end of software engineering" is about argument strength; this text, at least, does not declare the field over in so many words.