A Six‑Day, Million‑Token AI‑Driven Review Unpacks the L1‑L5 Agent Hierarchy

Paper production metrics

Using the DeliAutoResearch workflow together with DeepSeek‑V4‑Pro for writing and GPT‑Image2 for illustration, the authors generated a 46‑page, 538 KB LaTeX manuscript in six days. The process iterated six versions (V1 × 4, V2 × 1, V3 × 1), executed roughly 108 agent calls, consumed 648 000 tokens, and produced 2 234 lines of LaTeX source. The bibliography lists 103 verified references; the final paper contains 7 figures and 4 tables.

L1–L5 autonomy taxonomy for AI agents

L1 – Autocomplete : predicts the next token or line (e.g., early GitHub Copilot).

L2 – Task execution with human approval : agents decompose tasks but require human confirmation for each step (e.g., ChatGPT/Claude plus tool plugins).

L3 – Multi‑step execution with occasional human checks : agents autonomously perform 10–100 steps, requesting human review only at critical points (e.g., Claude Code, Cursor Agent).

L4 – Fully autonomous execution within a constrained domain : a human supplies only the research goal and final evaluation; the agent conducts experiments, writes code, and drafts papers but does not choose the research problem itself.

L5 – Fully self‑directed research : agents select topics, allocate resources, accumulate knowledge across domains, and conduct long‑term research. This level remains unrealized; the primary bottlenecks are continuous knowledge accumulation, trustworthy self‑assessment, and scalable architecture.

Four dominant architectural patterns

Single‑agent loop : a single model iterates reasoning → action → observation (e.g., ReAct, Reflexion, LATS, “thinking tree”). Simple and efficient but limited on complex tasks.

Multi‑agent collaboration : multiple specialized agents cooperate, providing diverse viewpoints and error correction (e.g., CAMEL, AutoGen, MetaGPT). Higher cost and coordination overhead.

Hierarchical scheduling : a top‑level planner delegates subtasks to lower‑level agents (e.g., Claude Code, Devin). Suited for long‑horizon, high‑complexity research; improves oversight.

Tool‑enhanced execution : agents interface with external tools such as code runners, browsers, APIs, or multimodal modules (e.g., SWE‑Agent). Performance depends directly on tool capabilities (Agent‑Computer Interface).

The authors emphasize that no pattern is universally superior; selection should match task characteristics such as length, complexity, need for multi‑view error correction, or external tool integration.

Six‑dimensional feature matrix evaluation of 17 autonomous‑research agents

The matrix assesses scalability, cost, reliability, autonomy level, domain specificity, and reproducibility. Results show a progression from fragile early prototypes to L4‑level, domain‑specific systems. Code‑focused agents rank highest in maturity, while scientific agents are beginning to produce verifiable discoveries.

Open research problems identified

Cognitive‑loop traps: agents may repeat ineffective strategies without self‑termination.

Context‑window limits: fixed windows (4 K–1 M tokens) hinder long‑term research.

Lack of automated novelty evaluation: no systematic method to assess originality or value of generated research.

Reproducibility challenges: model randomness and prompt sensitivity lead to non‑repeatable results.

Safety and ethics risks: dual‑use potential, autonomous self‑improvement, and academic integrity concerns.

High per‑task cost: single‑task expenses can reach ~50 ×  baseline, exacerbating research inequality.

Key bottlenecks for reaching L5 autonomy

The primary technical obstacles are sustained knowledge accumulation, reliable self‑evaluation, and scaling the underlying architecture to support continuous, cross‑domain research.

Code example

本文经AI新媒体量子位（公众号ID:qbitai ）授权转载，转载请联系出处
本文
约1500字
，建议阅读
5
分钟
本文介绍智能体 L1-L5 分级，梳理架构并剖析行业现存难题。