1 of 18

AI in Production

Aisha Yusaf

orra Co-Founder

Ezo Saleh

orra Co-Founder

Expertise in Applied AI and DeepTech since 2017. Tech veteran with over 20 years of experience at BBC R&D, JP Morgan Chase, and various startups. Specialises in optimizing complex systems, including distributed systems and Kubernetes tooling.

Launched an AI-powered children's camera game and a podcast audience insights platform. Graduate of Central Saint Martins and London Business School. Excels at achieving innovative solutions through multi-disciplinary collaboration.

2 of 18

WRANGLING WILD AGENTS

3 of 18

β€œIn March 2024 we built AI agents to scale buyer and seller onboarding for our marketplace startup.

We hit major production problems like service coordination, cascading errors and unnecessary agent calls racking up expensive LLM cost. With experience in distributed systems and developer tooling we decided to pivot.”

Firsthand AI Agent Challenges

4 of 18

Challenges to Deploying Agents

Latency

Slow workflows due to large data processing

Cost

Chatty agents driving up LLM costs

Reliability

Cascading errors making systems unpredictable

Integration

Difficulty integrating agents with existing infrastructure

5 of 18

Expectations vs Reality

Where 'just three agents talking' becomes a distributed systems PhD thesis

6 of 18

Key Realisations

Adaptive execution planning

Domain understanding

Balancing autonomy & control

7 of 18

The Missing Glue Layer

8 of 18

TECH DEEP DIVE

9 of 18

Monolithic Architecture

10 of 18

Monolithic Architecture

  • High Latency
  • Token Inefficiency
  • Reliability Issues
  • Debugging Complexity
  • Scalability Challenges
  • No Compensation Handling

πŸ‘Ž

11 of 18

Architecture Rethink

12 of 18

Architecture Rethink w/ orra

13 of 18

  • Reduced Latency
  • Lower Token Usage
  • Improved Maintainability
  • Better Reliability

How it works

  • Orchestration based on the work intent
  • Parallel Execution
  • Service Discovery
  • Seamless Integration
  • Health Monitoring

Multi-agent Architecture w/ Planning

πŸ‘

14 of 18

Multi-agent w/ Reliable Consistency

❌

  • Payment Gateway Down
  • Purchase Aborted
  • Inventory Stays Reserved
  • Products Appear Out of Stock

πŸ‘Ž

πŸ”₯

❌

15 of 18

Multi-agent w/ Reliable Consistency

  • Enable Consistent State
  • Automatic Recovery
  • Increased Reliability
  • Audit Trail

How it works

  • Compensation Framework
  • Automatic Triggering

πŸ‘

promote revertability

βœ…

16 of 18

Multi-agent w/ Reliable Plans

  • Hallucinates Execution Plans
  • Plans Don’t Match Intent
  • Invalid Action Sequences
  • Plan with Impossible State Transitions

πŸ‘Ž

‼️

‼️

17 of 18

Multi-agent w/ Reliable Plans

  • Ground in Desired Capabilities
  • Formal State Validation
  • Evaluate Pre-Execution
  • Don’t Execute Invalid Plans

How it works

  • Apply Domain Grounding in YAML
  • Embedding-based matching
  • Formal PDDL Plan Validation

πŸ‘

βœ…

18 of 18

Thank You!

Aisha Yusaf

orra Co-Founder

Ezo Saleh

orra Co-Founder