Stages of AI: thinking where we are heading in 2025
I think AI is going through the following stages: chatbots, assistants, co-pilots, agents or autobots
TL;DR: AI has rapidly evolved from basic chatbots to today’s 'agents' that can execute tasks. The post charts this progression towards a potential 'autobot' stage in 2025—fully autonomous AIs capable of independent, interactive decision-making—and highlights the crucial challenge of ensuring their actions align with human values.
The Evolution of AI Chatbots
Over the christmas break, I was vibe coding this site,varunsingh.net and based on the experience with pipecat.ai, I started to think about agents more concretely. I think AI is going through the following stages:
- chatbots (pre-2020, a chat interface responding to most common questions)
- assistants (2019 - soon replaced by agents, LLM-powered chatbots)
- co-pilots (2020 - human-in-the-loop, LLM-powered chatbots)
- agents (2023 - LLM with access to knowledge-base, APIs, databases, etc.)
- autobots or better name (2025 - agents that can take actions)
Understanding Each Stage
Chat bots are simple request and response bots that were rules-based, this was before we had LLMs.
An assistant is basically chatbots that were more reliable, similar to GPT 3.5/ChatGPT. These had inherent understanding of language and could string together compelling statements based on their training. With the help of Retrieval-Augmented Generation (RAGs) and vector databases, we are able to add use-case/customer specific knowledge-bases that the LLM can collate to form the response.
Co-pilots are as the word suggests assistants that have some persistence, i.e., either they are monitoring the actions that the user is making and based on those actions be able to provide guidance to the user. In coding, we have GitHub Copilot, Cursor, and other tools that are able to provide code suggestions based on the context of the code. In healthcare, there are several tools that doctors and medical providers are using to summarise patient notes, provide recommendations, and provide reminders. Lastly, customer support agents are getting pings from the AI co-pilot while they are conversing with the end-user or while they are working on a ticket.
Agents is the obvious next step, give the co-pilot or assitant the ability to take actions, i.e., the user of the Agent is moving from providing instructions to describing outcomes. This is big shift that we are seeing with code-generation, but can easily see this happening elsewhere like with Sales and CRMs, revenue recovery, simple support actions.
The Next Frontier: Autobots
Lastly, the autobots, I think some people call them auto agents, i.e., agents that can interact with other agents, take actions outside of their sandbox. In the above CRM example, we may have an artificial boundry that a CRM application may not automatically terminate an unpaid account with accrued dues of several months. In the agents stage, maybe it would send a notification to a human that the account is in revenue recovery for a few months, and delegate the decision of terminiation to the human in the loop, but in the autobot phase, it may decide on terminating access versus sending an extra set of reminder emails based on the value of the account. The thing we need to think about is how we ensure that autobots make decisions aligned with human values when they’re operating independently
Going into 2025, we are definitely in the Agents phase, the question is will we make autobots this year?