# Building AI Agents for Enterprise Operations at Parcha

## Company Overview

Parcha is building AI agents that automate operations and compliance workflows for enterprises, with a particular focus on fintech companies. Founded by AJ Asver (CEO) and his co-founder from Brex, the company started in March 2023 and has quickly moved from prototype to early customer deployments. The founders' experience at Brex, where they led the platform team responsible for automation, helped them identify the "last mile" of automation that previously required human expertise but could now be addressed with LLM-based agents.

## Technical Architecture and Implementation

### Agent Framework Evolution

- Initially built on LangChain for rapid prototyping and demos
- Moved away from LangChain's multiple abstraction layers for agent-specific code
- Rebuilt core agent codebase from scratch in ~500 lines of code for better reasoning and debugging
- Still uses LangChain for LLM interfacing and tool infrastructure
- Chose Claude (Anthropic) as primary LLM for speed and stability
### Agent Design Philosophy

- Breaks down complex workflows into smaller, manageable components
- Uses a hierarchical agent structure:
- Focuses on controlled, predictable behavior by using existing company procedures
- Avoids open-ended planning in favor of executing pre-defined workflows
### Prompt Engineering Approaches

- Uses context-setting to improve LLM performance
- Explicitly defines agent expertise and roles
- Structures agent plans as JSON lists with explicit step tracking
- Maintains careful balance of context window usage
- Avoids overloading context with too many scenarios
## Production Deployment Strategy

### Testing and Validation

- Requires 90% accuracy in back-testing before customer deployment
- Uses customer data for validation
- Implements sandbox testing environment
- Performs extensive back-testing with parallel agent execution
- Continuously collects customer feedback and edge cases
### User Interface and Integration

- Initially deployed as Chrome extension for transparency
- Shows step-by-step agent reasoning and decisions
- Plans to evolve into API endpoints for automation
- Focuses on building trust through visibility into agent decision-making
### Customer Empowerment

- Enables customers to modify and tune agent behavior
- Provides tools for benchmarking and evaluation
- Treats operations staff as future prompt engineers
- Allows real-time procedure updates without code changes
## Challenges and Solutions

### Technical Challenges

- Managing LLM hallucinations in compliance-critical contexts
- Scaling from simple demos to complex production workflows
- Handling edge cases in real-world scenarios
- Maintaining consistency across agent interactions
### Solutions Implemented

- Extensive white-glove customer support during early deployment
- Iterative improvement based on customer feedback
- Careful monitoring and logging of agent decisions
- Breaking down complex tasks into verifiable steps
## Lessons Learned and Best Practices

### Development Approach

- Rapid iteration over academic expertise
- Focus on velocity and learning from failures
- Importance of starting with constrained, well-defined problems
- Value of customer collaboration in development
### Agent Architecture

- Benefits of modular agent design
- Importance of transparent decision-making
- Need for careful prompt engineering
- Balance between automation and human oversight
## Future Directions

### Platform Evolution

- Moving toward self-service agent creation
- Building scalable enterprise platform
- Developing more sophisticated agent-human collaboration
- Expanding integration capabilities
### Industry Impact

- Trend toward hybrid workforces with AI agents
- Potential for increased business efficiency and scaling
- Evolution of emotional and collaborative agent capabilities
- Transformation of traditional operations roles