Summary of Key Points
This article focuses on the question of how companies can transform AI agents from demo versions into long-term, stable production systems that can operate efficiently. It covers various aspects, including the role of agents, technical architecture, product design, multi-agent systems, organizational adjustments, and talent requirements, addressing the critical issues involved in implementing agents within a company. Agents are not merely simple chatbots; they are intelligent assistants capable of helping companies complete actual tasks. To ensure their ongoing value, it is necessary to establish a stable technical framework, match them with appropriate business scenarios, and redefine the responsibilities of both humans and agents within the organization.
1. What exactly do agents do within a company?
Traditional enterprise software (such as CRM or ERP) serves as a “ledger” that records facts about customers, inventory, and finances. Agents, on the other hand, act as “intelligent assistants,” adding a layer of cognitive and operational capabilities to this basic functionality. For example, if you ask an agent to “analyze high-value sales leads from the past month and update them in the CRM,” it will understand the task, retrieve the necessary data, use the appropriate tools, execute the required actions, and then record the results directly in the system—without the need for you to open multiple software applications and copy and paste information manually.
The value of an agent is measured by two criteria: the number of tasks completed (e.g., handling 1,000 leads per month) and the value of those tasks (e.g., identifying 10 major customers). The number of conversations or the number of tools used are merely part of the process; what matters most to companies is the final outcome.
2. Why is AI coding the “testing ground” for agents?
In its early stages, agents were most successful in coding-related tasks because this context naturally lends itself to continuous model operations:
- Clear context: Code and documentation are stored in repositories that the models can directly access; moreover, the models have been trained on a large amount of GitHub code, making them familiar with this domain.
- Available tools: Tools such as terminals, compilers, and Git are readily available on computers, eliminating the need to rely on interfaces provided by business systems.
- Easy result verification: Whether the code is correct can be determined quickly through compilation and testing; if there are errors, they can be easily resolved using Git, resulting in low recovery costs (which is much better than sending out incorrect emails or making financial mistakes).
Therefore, coding agents serve as a pilot project for the development of more general-purpose agents, providing insights into their potential evolution.
3. When to use workflows and when to use agents?
Not all scenarios are suitable for agents; the decision depends on two factors: the level of business expertise required and the frequency of actions or reasoning processes:
- Use workflows for fixed processes: Tasks with fixed steps and stable rules, such as “sending notifications upon receiving forms” or “daily data aggregation,” can be more efficiently orchestrated using traditional workflow systems.
- For tasks that require specialized knowledge but few operations: For example, medical diagnoses, where expertise is needed but not much manual intervention, the focus should be on structuring business knowledge (using knowledge bases or graphs) to help models quickly find relevant information.
- Use general-purpose agents for complex tasks: Tasks like coding or in-depth research that involve multiple rounds of reasoning and action can be handled by agents.
- Use specialized agents for highly complex tasks: Tasks such as financial analysis or customer service require both specialized knowledge (e.g., understanding financial regulations) and multiple rounds of actions (e.g., checking customer history or generating solutions), which necessitate the combination of knowledge, processes, and tools.
4. What “core capabilities” are needed for production-level agents?
Agents in the demo phase can only chat and use tools; to be used in a production environment, they must possess the following additional capabilities:
- Harness (management layer): An assistant that prepares the necessary information (user tasks, historical data, tool instructions, permissions) before each model decision and ensures the task is executed smoothly. Many agent issues are not due to the models themselves but rather because the management layer has not provided the correct context (e.g., vague tool descriptions or outdated permissions).
- Tool library: A collection of specific actions that agents can perform, such as checking the weather, reading files, or accessing CRM interfaces. Tools should be designed with clarity (e.g., “check the weather in Shanghai”), structured parameters (e.g., only requiring the city name), and easy-to-understand results.
- Skill library: A set of instructions for how agents should perform tasks, such as “how to write a public account post,” which may include topic selection, structure, and formatting guidelines. These can be in the form of documents, scripts, or templates that models can use directly when needed.
- Stability measures:
- Transparency: Users should be able to see the tools used by agents and the actions they perform (sensitive information can be obscured).
- Human oversight: High-risk actions (e.g., sending emails or modifying data) should require user confirmation.
- Logging: All operations should be recorded for troubleshooting purposes.
- Progress and result visualization: Users should be able to directly monitor task progress and results without having to track every conversation.
5. How must organizations and talent adapt when implementing agents?
Agents are not meant to replace humans; they are designed to collaborate with them. Therefore, organizational and talent strategies need to change accordingly:
- Organizational adjustments:
- Merge job boundaries: Product managers may need to write code, and engineers should participate in requirement design.
- Establish AI-specific organizations that integrate data, processes, knowledge, and agent management (e.g., using a Context layer to connect all business data and a Pipeline layer to define task flows).
- Talent requirements:
- General skills: Quick learning ability, optimism about the potential of agents, resilience, and self-reflection for continuous improvement.
- Professional expertise: The ability to work from specific functions to entire product lines, with senior talents being able to identify opportunities in ambiguous areas.
- Business acumen: Understanding customers and the industry’s profit-making mechanisms (e.g., knowing how to allocate profits in advertising).
- Organizational skills: The ability to drive projects, coordinate resources, and resolve conflicts.
In summary, the successful implementation of production-level agents requires a combination of technical expertise, business understanding, and organizational adjustments. Start by identifying valuable business problems, build a stable system, and then consolidate these experiences into reusable capabilities to truly transform agents into productive tools for the company.