One Day, $1,000 in Tokens: An AI War Room for Legacy Code — Ford Prior, GEICO
Summary
This talk details how GEICO addressed the 'outer loop' bottleneck in their Software Development Lifecycle (SDLC)—the slow processes of decision-making, sign-offs, and context gathering—which historically limited the speed of coding (the 'inner loop'). They implemented a multiplayer, AI-native war room that uses agent fleets, structured rules (FleetKit), and human collaboration to accelerate major code modernization missions. The goal was to map a complex legacy claims domain end-to-end and deliver a functional change (like an agent-callable API) within one day and a limited token budget. The system successfully ran three modernization missions across rentals, damages, and salvage domains, producing fully vetted migration plans and design artifacts.
Key takeaways
-
The Outer Loop Bottleneck
In non-tech companies like GEICO, the SDLC is often limited by the 'outer loop' (decision-making, governance, context gathering), even if the 'inner loop' (coding) is fast. The solution requires automating the decision and intelligence layers.
-
AI-Native War Room Architecture
7:17
The war room is a collaborative system comprising key decision-makers, a 'driver,' and an AI-native productivity suite. It is wired into code repositories, workflows, agent skills, and enterprise MCPs to handle heavy lifting, fact-checking, and synthesis.
-
FleetKit: The Orchestration Layer
17:07
FleetKit is a structured collection of rules, protocols, and instructions written in markdown, optimized for guiding agent fleets (specifically for the Claude harness). It acts as a 'fleet manager,' taking a goal, breaking it down, dispatching research leaders, and managing the process through a human-in-the-loop queue.
-
Modernization Missions
11:43
The team ran three missions: 1) Uplift (rentals domain, extending an existing API to save costs); 2) Transformation (damages domain, building new agent-callable APIs); and 3) Reimagination (salvage domain, producing a fully vetted migration plan with API specs and design mock-ups).
Technical details
-
Agent Architecture
The system uses two types of agents: 'code dives' (for code crunching and graphing over legacy domain-specific languages in the cloud) and local agents (with access to local MCPs and enterprise-specific, highly optimized code).
-
Core Components
The architecture relies on a combination of vendor tools (e.g., Anthropic's code modernization plugin, the Claude harness) and internal GEICO assets (knowledge base, software standards, decision archaeology, and methods/skills).
-
Artifact Generation
The war room generates high-quality, structured artifacts, including HTML worksheets (using cloud artifacts), design documents (e.g., Draw.io or Figma visualizations), and complex visualizations of the architecture, which are crucial for human sign-off.
-
Model Selection
Benchmarks showed that models like Fable or Sonnet 5.5 performed comparably to, or better than, the previous generation of models when writing complex legacy code.
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.