Leveraging Generative AI Agents for Efficient Software Development

Agent Composition and Fundamentals

The Agentic Loop

• An agent runs tools in a loop while building context to achieve a goal
• Key components:
  • Language Model (LLM): Provides reasoning and draws on latent knowledge
  • Context Window: Stores working memory, system prompts, code, and history
  • Tool Selection and Execution: Uses MCP (Model-Centric Protocol) to interact with tools

LLM Capabilities

• LLMs provide reasoning and leverage latent knowledge from pre-training
• LLMs are stateless, so the context window is crucial for maintaining state

Managing the Context Window

• The context window stores all relevant information for the agent to operate
• Includes system prompts, context files, chat history, source code, and tool outputs
• Proper context management is essential for the agent to achieve its goals

MCP and Tool Integration

• MCP (Model-Centric Protocol) is the standard for integrating tools with agents
• MCP clients (e.g., Kuro CLI) can connect to various MCP servers to access skills, resources, and mutative actions

Kuro: Enabling Custom Agent Development

Kuro and Spec-Driven Development

• Kuro is a new agent-based AI development tool that follows a spec-driven development approach
• Provides a systematic way to define requirements, design, and break down tasks for software projects

Creating Custom Agents with Kuro CLI

• Custom agents are configurable AI assistants tailored to specific use cases and workflows
• Key components:
  • Model Selection: Choose the appropriate LLM (e.g., Haiku, Opus, Clots-on-it)
  • Prompt Definition: Describe the agent's personality and behavior
  • Pre-Approved Tools: Specify the MCP tools the agent can access
  • Resources and Hooks: Provide context and dynamic evaluation capabilities
  • Guardrails: Define security and compliance restrictions for the agent

Local vs. Global Agents

• Local agents are scoped to specific project directories and subdirectories
• Global agents can be accessed across multiple projects
• Use cases:
  • Local agents for project-specific workflows, personal productivity, and unique development environments
  • Global agents for general-purpose workflows, common development tools, and tool-specific workflows

Benefits of Custom Agents

• Workflow optimization by tailoring agents to specific tasks
• Reduced interruptions and better context awareness
• Improved team collaboration and security control

Bedrock Agent Core: Running Agents at Scale

Bedrock Agent Core Overview

• Comprehensive agentic platform that provides modular services for running agents at scale
• Includes capabilities like long-term memory, short-term memory, identity, and observability

FINRA's Custom Agent Journey

FINRA's AI Adoption Approach

• Rigorous R&D experimentation to establish trust and clarity
• Phased rollout with mandatory training and accountability
• Measurement program to track developer experience

Tech Upgrade Agent Use Case

• Challenge: Continuous technology stack upgrades competing with business priorities
• Solution: Developed a tech upgrade agent to automate the upgrade process
• Agent Features:
  • Scans code repositories and assesses upgrade complexity
  • Manages dependency version mapping and applies code changes
  • Validates the changes and issues a pull request for review
• Runs in both local and remote (Bedrock Agent Core) environments

Lessons Learned

• Importance of context engineering and management
• Applying clear boundaries and constraints as guardrails
• Continuous refinement based on pilot feedback
• Importance of cultural readiness and integration into workflows

Key Takeaways

• Agents can be composed of LLMs, context windows, and tool integration (MCP) to automate software development tasks
• Kuro enables the creation of custom agents tailored to specific use cases and workflows
• Local agents are project-specific, while global agents can be shared across projects
• Custom agents offer benefits like workflow optimization, reduced interruptions, and improved security control
• Bedrock Agent Core provides a platform for running agents at scale, as demonstrated by FINRA's tech upgrade agent use case
• Successful agent adoption requires careful context management, clear guardrails, continuous refinement, and cultural readiness