Why Making Autonomous Agents Useful Is So Hard?

Agent, the term became a hot topic in 2023—you've probably heard of AutoGPT, Stanford Town, BabyAGI, and countless other "LLM-based agents." These are not just concepts; they are real agents that can sense their environment, plan their next steps, and carry out tasks.

“That's definitely the future!” people were screaming. It was during this wave that PuppyAgent was founded. However, as people began to integrate these agents into their workflows, they immediately realized it was much more challenging than anticipated.

In short: Agent nowadays still aren't as useful as we thought they would be.

Why Making Autonomous Agents Useful Is So Hard?

Last year, we offered agents to our users and interviewed people around us. Surprisingly, even those most enthusiastic about agents gave a consistent answer: they aren't using "fully autonomous agents" in actual collaboration.

Decision-makers who buy in agent capabilities face a persistent concern:

Who bears responsibility if an agent's error leads to irreparable damage?

In enterprise settings, it's widely accepted that achieving both full autonomy and complete stability in agents is challenging. trading off between these two qualities is a crucial question that every company user has to answer. Our current real-world case suggest that companies tend to prioritize stability over autonomy.

Let's consider a hypothetical scenario: If an agent could save an employee 10 minutes daily but carried a 5% risk of causing a $5 million loss to the company, would you buy in it?

Two key questions:

1. In which case is the cost of an agent's mistake low?

2. In which case is the cost of verifying whether the agent made a mistake low?

By iterating on these two questions, we can identify potential SaaS for agents. RPA (Robotic Process Automation) and Flow. They are naturally suited for implementation because they either guarantee stability or limit the agent's capabilities through strict rules. This approach minimizes the cost of an agent's mistakes and improves process interpretability, making it possible to trace issues to specific steps in the process.

By the way, this also explains why over 50% of agent application scenarios are in RAG (Retrieval-Augmented Generation) Q&A and simple CRUD tasks. These scenarios sidestep the aforementioned critical issues.

In Which Way Can an Agent's Capabilities Be Used?

Only by addressing the two key questions mentioned above can an agent product truly find its PMF. Let's examine two examples:

Zapier (RPA Product)

Zapier, a typical RPA product, is pivoting towards agents. In April 2023, Zapier's agent debuted in OpenAI plugin store as a tool that generates fixed workflows from a single sentence (text to workflow). It also featured at OpenAI's DevDay in November 2023.

In Zapier, the LLM (Large Language Model) functions more as an RPA workflow-generating copilot than an agent. The LLM's value lies in simplifying RPA workflow creation, not in making real-time execution decisions. Most paying users seek solutions to problems that RPA can handle, aligning with Zapier's original positioning: "How to sync information from one software to another."

Many similar products exist, such as Make, which essentially extend RPA into the LLM era. They offer complete stability and error-free operation but lack autonomy.

Retool Workflow (Flow Product)

Retool, a well-established SaaS company, launched its workflow product in September 2023. This offering provides a no-code editing framework that focuses on low-code workflow editing and includes specific optimizations for RAG.

Similar products in this space include Dify, FastGPT, LangFlow, and Flowise. These tools enable quick and easy agent editing through interfaces, contrasting with RPA's rigid limitations on “agent behavior”. While they incorporate LLMs at certain stages, most logical decisions are handled by hard logic, minimizing the role of tool use. This approach ensures the stability of the agent's output through framework.

Flow-product are particularly well-suited for addressing corner cases within enterprises and align seamlessly with internal processes. Currently, the most widely adopted agent products in enterprise settings fall into this type.

What's PuppyAgent?

PuppyAgent is a startup founded in 2023,

We currently providing RAG services for enterprises, involving the build, deployment, and maintenance of a RAG system.

PuppyAgent's ultimate goal is to construct an agent-centric workspace that collaborates with humans.