Whitepaper "Agents" by Google (AI column)
I read it on the weekend. fancy whitepaper About agents from Google guys (Julia Wiesinger, Patrick Marlow, Vladimir Vuskovic). In this article, they describe a framework for creating AI agents that are capable of autonomous decision-making and interacting with the real world. To do this, they describe how such systems can integrate cognitive architecture with LLM models, thereby circumventing the limitations of the latter. This article reminds me of the Emerging Patterns in Building GenAI Products approach I’m talking about. told earlier. But the article "Emerging ..." is about the base. (Direct call LLM, prompt engineering, RAG)This one's about agents.
The authors specialize in AI infrastructure and cognitive architecture, or rather Julia Wiesinger on training agent systems, Patrick Marlow on orchestration instruments, Vladimir Vuskovic on decentralized AI systems. Their work is based on the development of LLM and agency workflow, and also presents Google’s Vertex AI as a foundational platform for deploying such systems, so it is actively used in all examples. In general, the guys at Google define agents as
An autonomous system that observes its environment, reasons about goals, and takes actions using tools to achieve outcomes It's important that the agent Monitors the environment through sensors Forms the internal state through reasoning Selects actions through a planner using tools (APIs, functions, databases)
This definition focuses not on static LLM models but on dynamic interaction with external data and tools. In general, architecture represents 1) The model LLM as a CPU for chain-of-thought reasoning 2) The tools Connects agents to the real world through three Extensions - API integration, execution on the agent's side Functions – generating code for execution that is executed on the client side, which causes the agent Data stores - work with databases that are usually vector and embeddings are used for search 3) The orchestration layer Management of iterative thinking using approaches of the type
- ReAct - Reasoning + Acting
- CoT - Chain of Thoughts
- ToT - Tree-of-Thoughts
The agents themselves can use it. three strategies for learning - In-context learningReal-time Adaptation Using Prompts and Examples - Retrieval-based learningDynamic access to external memory to make decisions with additional context - Fine-tuning: domain specific pre-training (pre-training) for optimization during use in this domain
Agents can work independently, proactively planning their steps to achieve goals without constantly engaging people. For example, they can prioritize tasks, adjust their strategy based on feedback from the environment, and recover from mistakes.
Google has an open source project.Project Oscar" to create agents for SDLC. It’s interesting to learn because it doesn’t focus on writing code. (Which is the fun part of this job.)And for the boring parts.
Oscar is a project aiming to improve open-source software development by creating automated help, or “agents,” for open-source maintenance. We believe there are many opportunities to reduce the amount of toil involved with maintaining open-source projects both large and small.
When it comes to the future of agents, Agents will independently manage computing resources, which can reduce dependence on centralized platforms. Agents will be able to solve more complex problems in collaboration Security issues will be important for agents to make it harder to exploit API vulnerabilities or gain unauthorized access to data in data stores.
As a result, this is a good basic whitepaper on the topic of building agent systems.
#AI #ML #Engineering #Software #Architecture #SystemDesign #DistributedSystems