AI Agent Evolution: From L0 Assistance to L5 Autonomy – Where Do You Stand?

AI Agent Capability Spectrum (L0‑L5)

To clarify the increasingly vague use of the term “AI Agent”, researchers and industry groups have borrowed the SAE J3016 vehicle‑automation grading system and defined six capability levels from L0 (no autonomy) to L5 (full autonomy).

Autonomy vs. Automation

Autonomy is not the same as automation. Traditional automation follows fixed scripts, while an autonomous agent makes dynamic decisions based on context, available tools, and observed outcomes. This distinction underpins all subsequent level definitions.

Level Definitions

L0 – No autonomy (pure human‑driven) : All tasks are performed by humans; AI does not participate. Typical scenarios include manual Excel reporting and hand‑written code.

L1 – Assistance (intelligent co‑pilot) : Stateless Q&A or suggestion generation; AI produces code snippets or answers, but humans retain final control. Example: GitHub Copilot code completion, early ChatGPT responses.

L2 – Partial autonomy (task‑execution expert) : AI can perceive environment, call external tools, and retain memory within a defined workflow, yet each step usually requires human confirmation. Example: ChatGPT or Claude that can search the web, read files, and call APIs.

L3 – Conditional autonomy (self‑executing agent) : AI can orchestrate end‑to‑end task flows and only asks for human input at critical nodes. Humans shift from operators to supervisors. Example: Claude Code’s “auto mode” (2026) that autonomously performs dozens to hundreds of programming steps, requesting review only at key points; Manus achieving leading scores on the GAIA benchmark; Lenovo LeXiang 4.0 enabling enterprise‑wide closed‑loop execution.

L4 – High autonomy (innovative explorer) : AI proactively discovers problems, proposes solutions, and executes them without real‑time human supervision. Example: an AI data analyst that automatically detects sales anomalies, traces causes, and generates a report.

L5 – Full autonomy (omnipotent creator) : AI can invent new methods and redefine tasks without any human involvement. As of 2026, no production‑grade system has reached this level.

Industry Landscape: Predominantly L1‑L2

According to iResearch’s 2026 “China Enterprise‑level AI Agent Development Insight Report”, most enterprises’ AI Agent maturity clusters around L1 (exploratory) and L2 (partial empowerment). The report notes that while adoption of large‑model‑centric AI is accelerating, depth of scenario implementation and realized business value remain below market expectations.

Adoption rates from the Shāqiú Think‑Tank survey show rapid growth: 17.3 % at the end of 2024, 25.4 % mid‑2025, and 40.3 % mid‑2026, yet most deployments stay at shallow L1‑L2 usage.

Emerging Shift Toward L3

Industry observers now agree the sector is entering the “L3 agent” era, a transition that began with the release of Claude Code in late 2025. Zhu Yibo, co‑founder and CTO of JieJieXing, declared at the 2026 Sullivan Summit that “the industry has undeniably entered the L3 era.”

Representative products accelerating this shift include:

Manus – the world’s first general‑purpose AI agent, leading on the GAIA benchmark.

Claude Code – dynamic workflow capabilities that can coordinate dozens to hundreds of agents for large‑scale tasks; its auto mode became the default paid‑plan setting in August 2026.

Lenovo LeXiang 4.0 – the first enterprise solution to move from question‑answering to closed‑loop execution.

Alibaba Qianwen – capable of autonomously operating a smartphone to place orders and book tickets.

These developments indicate a crossing from the “assistant era” (L1‑L2) to the “autonomous era” (L3).

Relation to OpenAI’s AGI Roadmap

OpenAI’s five‑stage AGI roadmap mirrors the L‑level framework: L1 (chatbot), L2 (reasoner), L3 (agent), L4 (innovator), L5 (organizer). The comparison highlights L3 as the pivotal jump from “thinking” to “acting”.

Challenges Ahead for L4 and L5

L4 requires agents to proactively discover problems and define tasks, shifting from reactive to proactive behavior. L5 demands the ability to invent new paradigms, a capability that remains at the research frontier and depends on breakthroughs beyond raw compute.

Conclusion

Most users today still operate AI tools at L1‑L2, where the AI assists but humans retain control. However, the opening of L3 – demonstrated by Claude Code’s auto mode, Manus’s autonomous task execution, and Lenovo’s closed‑loop AI – signals that AI agents are evolving from assistants to colleagues and, in some domains, to leaders. Understanding this spectrum helps individuals and organizations locate their current position and plan the next steps.