Huawei's 5-Level AI Agent Roadmap: From Tool Assist to Autonomous Wisdom (2035 Outlook)

The development of AI agents is a gradual evolution involving a comprehensive interplay among technology iteration, safety, business fit, and commercial ROI. Three core propositions must be resolved for large-scale social adoption: continuous capability improvement from tool to partner to independent decision-maker; a trust and responsibility framework defining human-machine accountability boundaries; and deep industry integration to unlock real value. This evolution correlates exponentially with market penetration: stronger agents enable broader scenarios, accelerating a positive value-chain loop.

The article references OpenAI's L1–L5 classification and presents Huawei's equivalent five-level taxonomy from the "Intelligent World 2035" report.

L1 Functional (Tool-assist)

Definition: Agent acts as a "conversational tool" executing single-step operations on command, with no autonomous planning.

Key capabilities: Shallow intent recognition, single tool/API invocation, short-term memory.

Typical scenarios: Retrieval QA, simple copy polishing, table formatting, single API calls.

Upgrade threshold: Introduce multi-step chain-of-thought and basic reflection to support at least two-step action sequences.

Metrics: Single-step task success rate, hallucination rate, response latency.

L2 Task Execution

Definition: Humans decompose and assign tasks; agent handles multi-step execution and result aggregation, still driven by external processes.

Key capabilities: Multi-tool orchestration, rule-based workflows, short-to-medium-term memory and state retention.

Typical scenarios: Report automation, customer-service ticket handling, RPA+LLM process robots.

Upgrade threshold: Enable agent to self-decompose/recompose tasks and adjust strategy during execution.

Metrics: Human intervention ratio, end-to-end completion rate, average task steps.

L3 Collaborative Autonomy

Definition: Agent self-plans and executes bounded-domain tasks; humans primarily collaborate and supervise.

Key capabilities: Long-term memory, plan-execute-reflect loop, multi-agent collaboration (division of labor/communication), environment interaction.

Typical scenarios: Marketing operations assistant.

Upgrade threshold: Cross-task, cross-system professional-grade decision-making and risk control, maintaining stable performance in complex uncertain environments.

Metrics: Unsupervised completion rate (unattended periods), cross-system call success rate, average interventions per complex task.

Industry consensus indicates most current deployments are transitioning from L2 to L3.

L4 Professional Guidance

Definition: Agent possesses expert-level guidance/service capability, autonomously deciding and continuously operating within clear boundaries; humans participate but do not over-intervene.

Key capabilities: High-fidelity world models/domain knowledge bases, long-horizon reasoning, safety audit and compliance policies, adaptive decomposition of complex goals.

Typical scenarios: Professional investment research assistant, clinical decision support, factory/park autonomous operations, L4 autonomous driving in designated scenarios.

Evolution judgment: Multiple sources project L4 agents will achieve scale deployment in certain industries/scenarios by 2035.

Metrics: Expert consistency (agreement rate with human experts), long-chain task success rate, risk-control/compliance pass rate, SLA attainment.

L5 Autonomous Wisdom

Definition: Full autonomy with cross-domain generalization approaching or surpassing humans; operates independently under an authorization framework and orchestrates other agents.

Key capabilities: Cross-domain transfer learning, meta-cognition (self-assessment/self-repair), swarm intelligence (agent-group co-evolution), robust value alignment and governance.

Typical scenarios: City-level autonomous operations, end-to-end unmanned manufacturing, self-balancing energy-compute-network systems.

Governance requirements: Strong-constraint ethical and safety governance (accountability, auditability, revocable authorization).

Metrics: Cross-domain zero-shot task success rate, autonomous duration (MTBI - mean time between interventions), system-level externality risk indicators.

Next Decade: 2035 Evolution Judgments

L4 Guidance Agents will scale in industries with mature engineering systems and clear rules (manufacturing, energy, telecommunications), where AI can match human work duration and intensity, enabling autonomous decision-making and full-lifecycle task execution.

L5 Wisdom Agents remain a longer-term goal requiring not only technical breakthroughs but also new legal, ethical, and social governance frameworks.

Value-chain acceleration: Agent proliferation will reshape production modes and business models across industries, evolving from point cost-reduction to full-chain value symbiosis.