AI agents are one of the hottest areas of development & investment in AI.
There’s no doubt that AI Agents can spark our imagination of what’s possible… building automation workflows, agents talking to agents, making autonomous decisions until the project is done. For example what if you could pay an AI agent to file your taxes, manage your stock portfolio, write your content, summarise your meetings and remind you of action items, and so on and so on. In practice however, AI agents are unfortunately still not so reliable, and you’d be taking a big risk by using them blindly.
Several venture capital funds including Insight Partners, Betaworks, Prosus Ventures and Activant Capital have created interesting content on the the topic of automation and AI agents, that I thought it’s worth a shot to try to synthesise it all into a blog post.
Here we go!
What is an AI Agent?
Agents are software systems designed to perform specific tasks or achieve predefined goals on behalf of users or organisations. Unlike traditional rule-based automation, agents leverage the latest advancements in natural language processing, machine learning, and decision-making algorithms to adapt to dynamic environments and execute tasks with context-awareness and nuance.
Agents can make decisions & take actions on their own following general instructions from a user.
Agents have 4 main components: – LLM – Tools: web search, code execution, etc – Memory: access to knowledge such as databases – Reflection & self-critiquing
- A powerful large language model (LLM) that understands user intent and creates an action plan based on the objective and the tools at its disposal.
- A suite of tools that extend the capabilities of the core LLM, such as web search, document retrieval, code execution, database integration, and other AI models, enabling the agent to execute actions like creating documents, running queries, or generating visualizations.
- Memory systems that provide access to relevant knowledge bases (long-term memory) and retain request-specific information across multiple steps required to complete an action plan (short-term memory).
- Reflection and self-critiquing mechanisms that allow more advanced agents to identify and correct mistakes made during the execution of their action plans and reprioritize steps as needed.
Difference between LLMs vs agents
Agents can come in various degrees of sophistication. This depends on the quantity and quality of the tools, the LLM used, constraints and controls placed on workflows created by the agents.
Agents can interact with users through natural language, interpret their requests, and then plan and execute a series of steps to achieve the desired outcome. This could involve tasks like scheduling appointments, performing research, automating business processes, or even providing personalised recommendations.
For further reading, there’s a whole body of works on going from RPA (Robotic Process Automation) to Agents.
The AI Agent Landscape
Prosus ventures created these two helpful landscapes on the current ecosystem for both AI Agent applications and the AgentOps ecosystem, a growing set of tools for creating, fine tuning and training new agents. Despite the growing number of companies in both categories, we’re still only in the early innings of smart AI automations. Sifted has a piece on what’s missing to make the promise of AI Agents a reality and I’ve listed some of the main limitations below.
The challenges of Agents today
While agents have the hype, they currently fail to deliver on expectations. Just search on Twitter for AutoGPT Failed, or BabyGPT failed and you’ll see plenty of example.
The three main challenges are:
- Technology readiness – building agents that can reliably understand context, make sound decisions, and execute tasks without errors or unintended consequences is no easy feat. There are many open ends and edge cases.
- Scalability of agentic systems (and cost) – agents often require significant computing resources and specialised tooling to operate effectively.
- Tooling and integrations – actions delegated to agents can be a bit of a black box. Ensuring transparency, explainability, and ethical behaviour in agents is crucial, particularly in sensitive domains like healthcare or finance.
AI Automation: Changing the Way We Work
An new report by Insight Partners covers various use cases for Generative AI Agentic use cases and made several predictions about the future of AI automation:
- Everyone will have an AI assistant: The integration of AI assistants into our daily workflows is becoming increasingly inevitable, enabling us to offload tasks and streamline processes.
- Human-in-the-loop is the operative framework: Deploying generative AI solutions will require a human-in-the-loop approach, where humans and AI systems work in tandem, leveraging each other’s strengths.
- Automation is a hard problem: Despite the hype, automation is often underestimated in terms of its complexity, requiring careful consideration and a gradual approach to deployment.
- Crawl, Walk, Run approach: The adoption of AI automation will follow a staged approach, starting with targeted use cases and gradually expanding to more complex scenarios as capabilities and trust increase.
- Code generation as a foundational element: The ability to generate code will become a foundational element of AI automation, enabling the rapid development and deployment of automated solutions.
Shameless Plug for Remagine Ventures
At Remagine Ventures, we recognise the profound impact that AI automation will have on the way we work, and we are monitoring the interaction between AI agents and the media, entertainment, gaming and commerce sectors. We’re excited both about the job-specific tools and the agentOps/ infrastructure to make Agents work for enterprise.
We learned that tools that target a specific job/role/ sector are likely to work better, and that the same risks that apply to generative AI investing (commoditisation, data privacy, price, etc) apply in the agent domain as well.
Israeli and UK pre-seed stage startups working on cool innovations in this space – we’d love to talk to you! You can find some other areas we are interested in, by reading the post on Remagine Ventures Requests for Startups.
I'm a former general partner at google ventures, head of Google for Entrepreneurs in Europe and founding head of Campus London, Google's first physical hub for startups.
I'm also the founder of Techbikers, a non-profit bringing together the startup ecosystem on cycling challenges in support of Room to Read. Since inception in 2012 we've built 11 schools and 50 libraries in the developing world.
- Israel’s Growing Generative AI Landscape – May 2024 Edition - May 14, 2024
- Everybody is talking about AI Agents and Automation - May 10, 2024
- Israeli Tech Scene Shows Glimmers of Recovery in Q1 2024 - April 30, 2024
