Core Characteristics of Agents and How to Navigate This Era

一. What Are the Core Capabilities of an Agent?

I believe there are nine key characteristics:

Decomposition as the Blueprint – The ability to effectively break down tasks.
Agents tackle task-oriented problems rather than providing simple answers, requiring iterative, multi-step solutions. Thus, effective task decomposition is the first fundamental feature.
Tools as the Means – The ability to utilize various tools.
Information and Context as the Basis – The ability to actively or passively acquire contextual information needed for task execution.
State Scheduling as the Engine – After decomposing a multi-step task, the Agent must efficiently manage and schedule subtasks.
- Analogy: An operating system abstracts running programs as processes, each with its own virtual address space, code, data, and execution state (register values, program counters, etc.). It encapsulates the minimal environment required for program execution.
- Similarly, an Agent abstracts the steps needed to complete a task as subtasks, each encapsulating specific goals, required inputs, outputs, execution logic (which may include tool calls), and its own state (pending, in progress, success, failure, etc.).
Reflection and Iteration as the Optimizer – The ability to self-reflect is a core capability of Agents.
I believe self-reflection is one of the most distinguishing features of Agents compared to LLMs. Google’s Deep Research shared a relevant diagram.
The challenge lies in how to reflect and how to evaluate the alignment between results and objectives.
Communication and Collaboration as the Bridge – The ability to collaborate with humans, clearly reporting task progress, decision rationale, encountered issues (especially when user input or authorization is needed), and execution results.
Agents solve tasks assigned by humans, but these tasks are often ambiguously described. Aligning the Agent’s actions with human needs is a critical capability—like a team member ensuring they’re always working on the right thing.
Contextual Adaptation as Flexibility – The ability to perceive and understand changes in the operating environment and dynamically adjust behavior.
Fault Tolerance and Robustness as Safeguards – The ability to handle inevitable errors and unexpected events during execution.
Continuous Learning and Evolution – The ability to improve across tasks over time.
“Continuous learning from experience, feedback, and new data to enhance long-term performance” is a core capability. I believe this is what distinguishes a basic Agent from a truly “intelligent,” evolving advanced Agent.
Recently, Manus’s open-source project released a reinforcement learning version, and MoonDark’s RL-based Agent achieved SOTA in the “Human’s Final Exam.” The field is clearly evolving in this direction.

II. How to Develop Key Skills in the Agent Era

As the Agent era approaches, what skills should we cultivate? I believe the core lies in insight, action, and management—with the biggest differentiator being management.

Skill 1: Deep Insight – Penetrate the surface to find the “anchor of demand” that AI cannot see.

Empathy will become even more critical in the AI era—understanding users, understanding oneself, observing one’s own needs, and effectively translating them into prompts.

Skill 2: Learning-by-Doing Action – Embrace “trial and error” to tame the AI beast through action.

Learn by doing: extensively experiment with AI tools, dive deep into their use, and move beyond mere testing.

Skill 3: A Manager’s Perspective – You are not AI’s rival but its “boss” and “conductor.”

Human value manifests in three dimensions:

Insight (Why): Use empathy to uncover the real, deep needs AI cannot reach, becoming the source of innovation.
Action (How): Start quickly, dare to experiment, and learn to wield tools through hands-on practice, turning uncertainty into fuel for growth.
Management (What): Shift your mindset—think like a CEO. Integrate AI and other resources to achieve goals, and take ownership of outcomes.

The AI era is about human-machine collaboration. The best strategy is to strengthen innate human strengths (insight and creativity), embrace the agility of action (learning and adaptation), and evolve managerial wisdom (integration and decision-making). These three pillars are our most solid competencies in the face of the surging AI wave.

III. Additional Thoughts on Agents

Among current Agents, my favorites are Google/OpenAI’s Deep Research. Google recently open-sourced the Gemini Fullstack LangGraph Quickstart framework, along with Bolt.new, Cursor, and other vibe-coding designs.
Among major companies, Tencent and Meta seem poised to benefit. Previously, they connected people; now, they can connect people-Agents and Agents-Agents, unlocking limitless possibilities.
NVIDIA is another likely winner—the Agent era will drive a surge in computing power demand. Unlike the past, where “thinking” consumed resources, now “action” will consume even more.

Let me know if you’d like any refinements!

Way to data scientist