Abstract
This survey investigates the intersection of software bots and cognitive architectures to identify "user-like" agents capable of true cognitive automation. By evaluating historical and modern AI agents against strict criteria—including environmental separation and recognized cognitive models—the study reveals a significant gap in current technology. Out of numerous systems reviewed, only one system from 1998 met all criteria for autonomous, human-like interaction with digital environments. The findings suggest that to build truly adaptable bots, developers must strictly separate the bot from its environment and integrate comprehensive cognitive architectures rather than relying on hardcoded API integrations.
Introduction
Software bots are ubiquitous in the modern digital landscape, ranging from Robotic Process Automation (RPA) in enterprise software to automated bug-fixers on platforms like GitHub. However, these bots are typically tightly coupled to specific platforms via Application Programming Interfaces (APIs). This coupling results in a lack of "situational awareness" and an inability to adapt to the dynamic, variable nature of the web.
When a human user navigates a new digital platform, they perceive "affordances"—such as buttons, links, and forms—and intuitively determine how to achieve their goals. Current bots generally lack this capability; they do not act like human users. This paper addresses this gap by investigating whether existing software bots possess "user-like" characteristics powered by a cognitive architecture: a holistic computational model of human-like thinking that encompasses memory, perception, and decision-making. The ultimate objective is to understand how to engineer bots for true cognitive automation, enabling agents to autonomously perform complex knowledge work across diverse digital environments.
Methodology
The authors conducted a systematic literature review using forward and backward snowball sampling to identify research claiming to develop bots with cognitive capabilities. Starting with an initial pool of approximately 190 works, the field was narrowed to 11 representative software bots and agents spanning the years 1996 to 2018.
These 11 candidates were evaluated against four strict criteria:
- Software Bot: The system must automate a task, initiate actions proactively (rather than just responding to prompts), and operate externally to the target environment.
- User Similarity: The bot must act like a human user, maintaining a clear separation from the environment it is manipulating.
- Cognitive Architecture: The bot must implement a recognized cognitive model (e.g., LIDA, ACT-R).
- Cognitive Abilities: The bot must exhibit high-level generalized cognitive functions, excluding systems that rely solely on narrow, low-level heuristics.
Analysis and Key Results
Because this is a literature survey, the "data" consists of the 11 selected historical and contemporary software agent systems, such as CALO, IDA, and World of Bits. The evaluation yielded a surprising result: out of the 11 representative systems, only one met all four criteria.
The Success of IDA
The sole system to pass all tests was IDA (Intelligent Distribution Agent), developed by Franklin et al. in 1998 for the US Navy. IDA was designed to assign tasks to sailors via email, interacting with the environment in a way that mimicked human decision-making and communication protocols of the time.
Reasons for Failure in Other Systems
Most other systems failed the "User Similarity" criterion. The survey noted a distinct shift in software engineering after the year 2000; developers increasingly built bots directly into the systems they were automating. While this direct integration improves efficiency for specific tasks, it sacrifices the user-like independence required for generalized adaptability.
Furthermore, modern attempts often fall short on high-level cognition. For example, the "World of Bits" (2017) platform trains bots to use actual mouse and keyboard actions on web Document Object Models (DOMs). While this satisfies the requirement for user similarity, these agents typically rely on low-level reinforcement learning operations and lack the high-level cognitive architectures (such as self-awareness or contextual reasoning) required to pass the cognitive abilities test.
Discussion: The "User-Like" Gap
The survey highlights a significant missing link in modern AI. Despite the proliferation of conversational AI, most current systems act as interfaces for users rather than acting as autonomous users themselves. To achieve generalizable intelligence, the authors argue that bots must be decoupled from their host environments. By treating a bot as an external entity that interacts with a User Interface (UI)—much like a human switching between browser tabs—developers can achieve higher levels of adaptability.
This requires a paradigm shift in software engineering. Developers need to componentize cognitive architectures so that high-level behavior models can be dynamically plugged into software bots, allowing for self-awareness and contextual awareness in digital tasks.
Conclusion and Future Work
The study acknowledges limitations, including a focus on publicly available literature which may exclude proprietary enterprise bots. However, the results clearly indicate that the path to autonomous digital workers lies in the practical integration of cognitive architectures into modern bot development. Future work should focus on refining these evaluation criteria and expanding the search to include emerging cyber-physical systems and smart factory environments, where autonomous navigation of complex interfaces is increasingly critical.
Summary updated