One Day, $1,000 in Tokens: An AI War Room for Legacy Code — Ford Prior, GEICO

Highlights

• Engineers at GEICO spent only 25 percent or less of their bandwidth on major competitive changes before implementing managed AI agents.

• A multiplayer AI-native war room running for one day with $1,000 in tokens successfully mapped out and modernized legacy domains end-to-end.

• The rentals domain modernization experiment used an agent to discover a connection that saved 8,500 calls per month.

• Fleet kit utilizes a tight structured collection of rules, protocols, and markdown instructions to direct architecture transformation tasks.

• Model testing revealed that Sonnet 5.5 wrote code just as well as, if not better than, Fable.

Timeline

1. 00:00:00 - 01:34:00

   Software Development Life Cycle Bottlenecks

   While automated coding tools have sped up the build process, organizations remain rate-limited by collaboration and intelligence gaps during major changes. GEICO addressed this by deploying managed AI agents across various harnesses to shift engineer bandwidth toward competitive development work.

   • Inner loop coding speeds have accelerated significantly while outer loop processes remain constrained.
   • Engineers dedicated 25 percent or less of their bandwidth to major competitive version changes.
   • Strategic investments in managed AI agents aim to offload patches and minor work from spec to deployment.
2. 01:34:00 - 06:25:00

   The Multiplayer AI-Native War Room

   To overcome collaboration and intelligence bottlenecks, an AI-native war room was established to synthesize variables, fact-check, and crunch code. The primary challenge tested whether a domain could be mapped end-to-end and deployed as an agent-callable API within a single day.

   • A multiplayer AI-native war room combines key decision-makers, subject matter experts, a tool driver, and a productivity suite.
   • An experiment tested whether one day and $1,000 in tokens could map a legacy domain and propose a serious change.
   • Legacy claims systems containing millions of lines of code served as the primary testbed for domain modernization.
3. 06:25:00 - 08:26:00

   Modernization Experiments across Domains

   Three code modernization missions tested the approach across increasingly complex domains. These missions ranged from uplifting rental vehicle APIs to transforming damages data workflows and fully reimagining the salvage domain with ratified project plans.

   • The rentals domain experiment utilized an agent to uncover data connections that saved 8,500 calls per month.
   • The damages domain focused on building new agent-callable APIs to unlock larger workflows.
   • The salvage domain produced a fully vetted migration plan and design mock-ups greenlit by product and tech leadership.
4. 08:26:00 - 16:00:00

   Architecture and Fleet Kit Operations

   The operating model relied on a division between vendor-provided infrastructure and internal GEICO assets like knowledge bases and software standards. A structured fleet kit framework steered agents to execute tasks methodically, producing artifacts that teams could review and edit.

   • Fleet kit consists of structured rules, protocols, and instructions written in markdown optimized for the Claude harness.
   • The fleet manager breaks down goals, dispatches research leaders, and passes work through a human-in-the-loop approval queue.
   • War room tools generate high-quality HTML worksheets and complex architectural visualizations.
5. 16:00:00 - 01:34:34

   Lessons Learned and Future Directions

   Running autonomous agents at scale yielded critical lessons regarding prompt caching, budget management, and model selection. Moving forward, the focus shifts toward automated meeting transcription, better context handling, and scaling agent deployments across the organization.

   • Improper cron job configurations in AI sessions can rapidly deplete token budgets.
   • Sonnet 5.5 matched or exceeded the coding performance of Fable during benchmark tests.
   • Future enhancements include improved gate enforcement, optimized context search, and automated audio transcription.