# Building Jarvis: JPMC’s Agent Building Assembly Line

## Executive summary

JPMorgan Chase introduced 'Jarvis,' an agent-building assembly line, which fundamentally shifts the process of developing complex AI agents. The system standardizes agent creation, reducing development time from months to minutes. Jarvis is grounded on three principles: speed, standardization, and governance. It manages the entire lifecycle—from Software Development Life Cycle (SDLC) to Agentic Development Life Cycle (ADLC) and Production Development Life Cycle (PDLC)—by making all components (skills, agents, UIs) composable, while maintaining rigorous governance through continuous offline and online evaluations.

## Key takeaways

- From Months to Minutes: The team demonstrated that the time required to build a complex agent has been drastically reduced from months to minutes by implementing a standardized assembly line approach.
- The Assembly Line Model: Jarvis treats agent development not as custom software, but as a composable assembly line, ensuring standardization and speed across the entire development process.
- Three Pillars of Jarvis: The assembly line is built on three principles: 1) Speed (concept to production in days), 2) Standardization (one common framework), and 3) Evaluations and Governance (incorporating governance early).
- Continuous Trust and Governance: Trust is maintained through two continuous disciplines: Offline evals (using positive and negative test cases) and Online evals (real-time monitoring for hallucination and relevancy checks).
- The Human Core: Human Subject Matter Experts (SMEs) remain central. They are promoted to 'agent team leads' to provide crucial orchestration and collaboration skills, ensuring the team works cohesively.

## Technical details

- Agent Development Life Cycles: The system covers three distinct life cycles: Software Development Life Cycle (SDLC), Agentic Development Life Cycle (ADLC), and Production Development Life Cycle (PDLC).
- System Architecture and Composability: The design emphasizes composability, allowing skills, agents, and UIs to be replaced or swapped out (e.g., switching between private/public cloud or different deep agent stacks).
- Enterprise Integration and Security: Implementation requires an enterprise-level harness that addresses security, memory management (long/short memory), inter-agent communication, and sandboxing.
- Evaluation Frameworks: Evaluation includes: 1) Offline evals (defining positive and negative boundaries); 2) Online evals (measuring hallucination rate and relevancy check in real-time); 3) Ongoing Performance Monitoring (OPM).
- Skill Abstraction: The system differentiates between reusable 'engineering skills' (the common pipeline) and unique 'business skills' (the domain-specific differentiator provided by human expertise).

## Practical implications

- Build teams can adopt a standardized, assembly-line approach to agent development, significantly accelerating time-to-market.
- Focus on building composable systems ensures flexibility across different cloud environments and integration protocols.
- Governance must be continuous, integrating human review and automated evaluation (hallucination/relevancy) from the first answer generated.
- The role of the Subject Matter Expert (SME) shifts from merely providing knowledge to providing critical 'orchestration' and 'collaboration' skills to agent teams.

## Topics

Agentic Development, AI Governance, CI/CD, Composability, Financial Technology, LLMs, Deep Agents, Offline and Online Evals, Interrupt, the agent conference by LangChain

Source: https://www.youtube.com/watch?v=VpHYPfvfy0o
