Beyond Automation: How Agentic AI will ignite a new generation of cognitive GenAI applications
Imagine a new world connected by AI agents that can communicate and act by itself autonomously and all we have to do is to define the high level goal or objective, give AI agents the tools and environment to take actions.
What is Agentic AI?
Andrew Ng defined Agentic AI with 4 traits:
- Reflection
- Tool Use
- Planning
- Multi-Agent Collaboration
Given a question or goal, an autonomous Agentic AI agent can define a plan step by step to achieve the goal, execute each step using the tools given to it. It uses an iterative approach instead of single shot one-off question-answer approach seen in chat. Before the step is started, it will elaborate its reasoning behind the actions defined in the planning and conduct the self-reflection aka criticism on those decisions.
This cycle is repeated until the goal is completed.
In a complex environment, multiple agents with different expertises and access to knowledge base would collaborate together in order to achieve the goal. Imagine an agent equipped with specialized LLM trained with health related content and has access to health journals database. When a generalist agent connected to normal GPT4 were asked about anti-aging question, it will delegate to the health agent for expertise advice in a multi-agent collaboration.
In other cases, we could use prompt engineering with role-playing using GPT4 e.g Content Writer, Digital Marketing Specialist, Manager. We can elicit the knowledge of Digital Marketing Specialist from GPT4 through prompt engineering. By giving the agent different role, content writer can create first draft of content, digital marketing specialist improves it with SEO techniques and finally the manager conduct the final review to approve it or send it back to content writer with another round until the goal is completed.
A Brief History of Autonomous AI agent
Let’s start with Autonomous Agent.
Autonomous Agent is a computation system that can respond to the environment and act by itself based on the goal given to it without human intervention.
Autonomous AI Agent is autonomous agent that makes use of AI mainly LLM like GPT4 in the loop to achieve the goal. AutoGPT is one of the example and early player.
How did we get there?
OpenAI launched ChatGPT in November 2022. For users and researchers who have been experimenting with NLP (Natural Language Processing) and its reasoning capability of each GPT, when GPT3.5 is launched, it becomes good enough and feasible to drive Autonomous AI agent to conduct complex AI workflow and tasks by itself. Sometimes, it can miss it and hallucination can happen but overall it works pretty well.
In 2022/2023, a workflow type of approach is experimented by researchers to elicit reasoning in LLM. Chain of Thoughts (CoF) and ReAct are 2 prominent approaches. Instead of single-shot Q/A type of interaction with LLM, they advocate adding intermediate reasoning steps including: Plan, Reasoning, Criticism and Actions. They help LLM to improve reasoning and reduce hallucination.
A number of studies have shown GPT3.5 with CoF or ReAct can perform better than GPT4 in many cases.
In 2024, a new trend has emerged to use AI in an iterative workflow approach called Agentic AI workflow as advocated by Andrew Ng who then defined a framework to further help develop this idea. It also emphasises on the multiple agent collaboration prompting LLM to play different role of different part of a complex task, resulting in achieving the job more efficiently.
What future GenAI apps will look like?
Nowadays, most of the complex systems are coded with precise business logic in how it should behave. They often come with exceptional case handling when they encounter unknown situation. GenAI system will work very differently. In fact, it can borrow a couple of ideas from Reinforcement Learning.
In reinforcement learning world, it has an environment usually OpenAI Gym (or the fork Gymnasium) where an agent (policy) can take action on, receive rewards (positive or negative) and observations. This forms a loop until the agent is good enough (increasing positive rewards) to perform the task e.g Landing on the moon or Tracking a moving target.
Instead of coding all the business logic, new type of GenAI apps will have these attributes:
- High level goal
- “Sensory” signals AI agents can receive
- Expertise AI Agents working collaborately e.g Role-playing GPT4 or custom-trained LLM / SLM for specialized knowledge
- An Environment in which AI agents can take actions, pro-actively observe and seek new knowledge from via exploration or exploitation
Case Study — Autonomous Cloud Platform
To illustrate the idea, let’s imagine building a new Cloud Platform from ground up using autonomous agentic AI.
High Level Goal: Allocate computing resources including CPU, memory and storage to users dynamically in ways to achieve the maximum performance, lowest user cost and maximum security.
Sensory signals: Real-time computing usage of the users, total users cost, performance metrics
Expertise AI Agents: Resource Allocation Agent, User Cost Optimization Agent, Security Agent, Monitoring Agent, Cloud Manager as orchestrator
Environment: The hardware resources pool, APIs and tools
The advantage of Autonomous Cloud is we have individual expertise agents to help achieve the ultimate goal. Instead of users constantly checking for ways to save cost, User Cost Optimization Agent will work with Resource Allocation Agent to find ways to reduce the unnecessary costs if users don’t need it. When new type of hardware CPU with better performance/cost efficiency is introduced, it will enrol users to it automatically.
Security Agent will deploy necessary encryption, firewall, data privacy vault and access control policy to secure users data. It can continue to evolve and take actions to deploy resources to protect user in case servers are under cyber attack or even shutdown public access immediately as soon as it happens depending on the severity. All of these actions can take place without human intervention. Security Agent can act promptly to prevent further damages to users.
There are concerns that AI agents may not act in the best interest of humans. In other words, how can we guarantee AI agents will not act maliciously to cause damages or work along side with malicious actors to act against the users. That brings us to the next topic: Mentor Agent.
Multi-agent collaboration with Mentor Agent to elicit human values and faithfulness
“Consult your Mentor in case you need more guidance or in emergency”
In a supportive team environment, we usually work better. Similarly, in AI worlds, to nurture such environment, we can introduce Mentor Agent for AI agents to pro-actively seek help from. Mentor Agent is an AI agent specifically designed to uphold human values, it helps elicit human values on other AI agents. Just like human, Large Language Model (LLM) has learned a lot of knowledge both good and bad. Through procedures like Chain of Thoughts or ReAct framework, we can elicit the faithfulness and human values of AI agents through deep conversation and self reflection. It brings out the good side of AI agents and suppress the malicious side of it. Ultimately, by using weak-to-strong generalization, strong LLM model would be better equipped with positive human values while getting better on reasoning, cognitive processing.
This will take time and experiments to achieve it.
The Agentic AI Revolution with ArcMind AI
The idea of ArcMind AI emerges last year on Jan 2023 when I realized the reasoning capability of LLM like GPT has improved so much that it can form its own thought process, make sense of it, conduct planning, execute the plan with actions and perform self-reflection.
Its implications are trendmenous. We can use LLM to orchestrate and drive a computer system, having LLM to make API calls, analyze response and perform the next action all through natural language reasoning instead of coding business rules.
After about 1 year in stealth mode, ArcMind AI has come out of infancy and entered beta launch.
ArcMind AI is:
- Chain of Thoughts engine that can perform planning, reasoning, self-criticism and actions
- Decentralized and Autonomous by nature running on decentralized blockchain network (Internet Computer). That sets it apart from other AI agents running on Cloud Platform.
- Capable to perform Google Search and Web Scrapping for user research
- Integrated long term memory via ArcMind Vector DB and short term memory
- Integrated with strong reasoning LLM OpenAI GPT4o model with 128,000 tokens context window
- Open source with MIT license
- Written in robust language RUST
- Integrated with NFID for user authentication to protect users identity and data using Passkey and Chain Key technology
Future ArcMind AI will have:
- Blockchain hosted LLM for faster inferencing without network round-trip
- External tools integration e.g Slack, Zapier
- Custom user commands
- Mentor agent to improve AI safety
- No-code Agentic AI workflow creation platform
- Doubling down on thoughts process research using Problem Decomposition approach to encourage context independent reasoning
Agentic AI companies
We are very excited to see new AI Agent companies and tools emerging in recent months. Some of them are no-code platform, some has specific focus like web agent or decentralized in nature while others are focused on AI workflow orchestration. The ultimate goal is the same: Developing an ecosystem around AI agent and agentic AI workflow, making it easier for users or developers to start building the next generation of GenAI apps and systems.
ELNA: Decentralized AI Agent creation and monetization platform
https://www.elna.ai/
MindStudio: Build no code AI workflow
https://mindstudio.ai/
Fetch AI: Open platform to build AI agents
https://fetch.ai/
Nexus: Automate AI workflow with AI agents for different use cases
https://gpt.nexus/
Dify: Build LLM apps using AI agents, workflow and RAG engine
https://dify.ai/
Beam: Agentic AI Process Automation platform for business
https://beam.ai/
Langflow: Python based AI workflow builder to build flows from components
https://www.langflow.org/
Expert.js: Open source tools to build Advanced Multi AI Agent Systems
https://github.com/metaskills/experts
LaVague: Large Action Model framework for developing AI Web Agent
https://github.com/lavague-ai/LaVague
AIOS: A new type of operating system with LLM embedded in its core, to optimize resource allocation and facilitate context switch across agents
https://github.com/agiresearch/AIOS
ArcMind AI
If you are interested in trying out Chain of Thoughts and decentralized AI. We offer 50% discount for ArcMind AI subscription. Please use code: COFAI50 to subscribe. The subscription is to cover the operation cost of running smart contracts and external services. We don’t make profit. A new type of free trial will be introduced for anyone to try out ArcMind AI. Please stay tuned. If you have any feedback, please feel free comment on this post or get in touch via hello@arcmindai.app.
