5 AI Agents
Traditional AI interactions follow a simple pattern: you ask a question, the system provides an answer, and the conversation ends. While this approach works well for straightforward queries, legal practice often requires complex, multi-step processes that would benefit from more sophisticated AI assistance. The more complicated the task or project, the more traditional prompt-and-response interactions would require dozens of separate queries, manual coordination between steps, and constant human intervention to maintain context and momentum.
AI agents offer a fundamentally different approach: autonomous systems capable of planning, executing, and adapting complex workflows with minimal human oversight.
What Are AI Agents?
AI agents are autonomous or semi-autonomous systems that operate without human supervision. Agents can perceive their environment, make decisions, and take action to achieve specific goals. These systems typically combine LLMs with planning capabilities, memory components, and tool usage to perform complex tasks through reasoning, planning, and execution.
The model acts as the planner, serving as the decision-maker for the agent’s processes. As previously discussed, LLMs are limited by their training and do not engage with the outside world. Tools allow agents to access and process external information. The search feature on ChatGPT is an example of a tool. When enabled, ChatGPT searches the web to assist in answering a prompted question.
An agent’s memory components enable it to recall how it has tackled tasks in the past or how it has interacted with users, allowing for more personalised future experiences.
Intelligent personal voice assistants, such as Siri, Alexa, and Google Assistant, are excellent examples of agents. These agents assist with various tasks upon command, such as scheduling events, setting reminders, sending messages, and searching for information.
Overall, the workflow of an AI agent involves reasoning, planning, and execution.[1] When a user prompts the agent, it will:
- Receive the command,
- Process the request, figure out how to approach it, and
- Get the task done.
Unlike traditional AI applications that respond to direct commands, agents can maintain context across multiple interactions, adapt their strategies based on feedback, and proactively work toward achieving their objectives.
An example of an available AI Agent is Spellbook’s Associate. The AI agent is marketed as being able to work through multi-document legal matters, completing tasks for you across multiple documents.[2] Prompting the agent enables the updating or fixing of contractual terms and party details across multiple document sets.
Watch
Watch this video where Maya Murad explores the evolution of AI agents and their pivotal role in AI systems.
Core Components of AI Agents
AI agents typically comprise four essential components that work together to enable autonomous operation:
1. Planning Engine (The Strategic Mind) The planning component serves as the agent’s strategic intelligence, responsible for understanding complex objectives and decomposing them into executable subtasks. When you assign a complex legal task, the planning engine analyses the requirements, identifies necessary steps, sequences activities logically, and creates contingency plans for potential obstacles.
For example, when tasked with “preparing a comprehensive intellectual property audit for our client’s acquisition target,” the planning engine might develop a strategy involving:
- Document collection and organisation
- Patent portfolio analysis
- Trademark registration verification
- Copyright and trade secret identification
- Risk assessment and reporting
- Stakeholder communication planning
2. Memory Systems (The Institutional Knowledge) Memory components enable agents to maintain context across extended workflows and learn from previous experiences. This isn’t simply conversation history—it’s structured knowledge management that includes:
- Working Memory: Current task context, intermediate results, and temporary findings
- Episodic Memory: Detailed records of previous similar tasks and their outcomes
- Semantic Memory: Accumulated knowledge about legal principles, client preferences, and successful strategies
- Procedural Memory: Learned workflows and optimised approaches for everyday legal tasks
3. Tool Integration (The Practical Capabilities) Tools represent the agent’s ability to interact with external systems and perform concrete actions. In legal contexts, these might include:
- Research Tools: Access to legal databases, case law repositories, and regulatory resources
- Document Tools: Ability to read, analyse, create, and modify legal documents
- Communication Tools: Email systems, calendaring, and client management platforms
- Analysis Tools: Financial modelling, risk assessment, and compliance checking systems
- Data Tools: Database access, information extraction, and reporting capabilities
4. Execution Engine (The Operational Controller) The execution component coordinates all other elements, manages workflow progression, handles error recovery, and maintains overall system coherence. It ensures that planned activities are executed efficiently, monitors progress against objectives, and adapts to changing circumstances.
How AI Agents Differ from Other AI Approaches
Understanding the distinctions between different AI approaches enables legal practitioners to select the most suitable tool for specific tasks.
Direct LLM Interaction
Human Input → LLM Processing → Single Response → End
Characteristics:
- Immediate response to specific queries
- No memory between interactions
- Limited to information in the training data
- Requires human orchestration for complex tasks
Best For: Quick questions, draft generation, brainstorming, simple analysis
RAG-Enhanced Systems
Human Input → Knowledge Base Search → LLM Processing (with retrieved context) → Response → End
Characteristics:
- Access to current, specific information
- Enhanced accuracy through grounded sources
- Still requires human orchestration for multi-step tasks
- Limited autonomy in information gathering
Best For: Research queries, document analysis, and current information needs
AI Agents
Human Objective → Planning* → Tool Selection → Execution → Adaptation* → Result → Memory Update
* A feedback loop occurs between these stages
Characteristics:
- Autonomous multi-step execution
- Persistent memory across tasks
- Dynamic tool selection and adaptation
- Continuous learning from experience
- Complex workflow orchestration
Practical Example: Contract Portfolio Review
Direct LLM Approach
You: “Review this employment contract for compliance issues.”
AI: Provides analysis of a single contract based on general knowledge.
Result: Basic analysis, no comparison across portfolios, no current law.
RAG-Enhanced Approach:
You: “Review this employment contract for compliance issues using current Australian employment law.”
AI: Searches the current employment law database, analyses a single contract with the current authorities.
Result: Current, accurate analysis of a single contract.
AI Agent Approach:
You: “Conduct a comprehensive compliance review of our employment contract portfolio.”
Agent:
- Inventories all employment contracts in the system
- Identifies relevant compliance frameworks for each contract type
- Research current legal requirements and recent changes
- Analyses each contract against applicable standards
- Identifies patterns of non-compliance across the portfolio
- Prioritises issues by risk level and urgency
- Generates a comprehensive report with specific recommendations
- Schedules follow-up reviews for time-sensitive matters
Result: Complete portfolio analysis with actionable insights.
Lessons on AI agents from Claude Plays Pokémon
To give you an example of what AI agents can look like, watch this video where Alex Albert and David Hershey discuss the story behind Claude Plays Pokémon. This experiment demonstrates how AI agents can navigate complex tasks.
Explore
For more information about AI Agents, check out the white paper from Google on AI Agents.
