GEICO sped up coding with AI agents, but decisions, approvals, context gathering, and collaboration still slowed major software changes.
2
Ford Prior tested whether $1,000 in AI tokens and one day could map a claims domain, propose an agent-callable API, and produce a working demonstration.
3
FleetKit coordinates research agents, human approval queues, worksheets, design documents, and builder agents around GEICO's legacy systems.
Summary
Ford Prior describes how GEICO moved its AI effort beyond code generation. The company's inner software loop became faster, but the outer loop of decisions, approvals, domain research, and collaboration remained slow. His answer was a multiplayer AI-native war room with decision makers, subject matter experts, one driver, connected tools, and agent fleets. Prior ran a one-day, $1,000-token experiment three times across rental, damage, and salvage claims domains. The work produced an extended rental API, new damage APIs, and a vetted salvage migration plan with product and technical approval. FleetKit, a markdown rulebook for Claude-based agent fleets, breaks a mission into research, human review, and implementation. Agents also produced HTML worksheets and architecture diagrams that teams could edit and reuse. Prior is candid about wasted tokens, prompt caching, model selection, and the limits of laptop-based orchestration. He wants stronger gate enforcement, better context search, audio capture for meetings, and wider organizational deployment.
GEICO sped up coding while the rest of software delivery stayed slow
Prior says GEICO had largely removed the large coding and intellectual-work bottleneck from its inner loop. The outer loop still included everything around coding and building, including risk management, decisions, approvals, and collaboration. GEICO is a risk management company, so those controls are part of how it develops and delivers software. Engineers also spent much of their time on patches and minor upgrades. After those tasks, they had 25% of their bandwidth or less on a good day for major changes from one semantic version to the next.
The war room combines people, tools, and agents around major decisions
The proposed solution is a multiplayer AI-native war room. It brings together decision makers and subject matter experts, with one driver operating the expensive AI tools. The setup includes communication channels, an interactive design studio, code repositories, workflows, agent skills, enterprise MCPs, and agent fleets. Prior says the missing capabilities were collaboration and intelligence: the ability to bring the right people together and connect code, facts, and other variables quickly enough to support decisions.
The experiment aimed to map a legacy domain and ship an API in one day
Prior set himself a concrete target for GEICO's claims systems, which contain millions of lines of code. With $1,000 in AI tokens over one day, he wanted to map one domain end to end, including its business topology and legacy code, then have the agents propose an agent-callable API. The final target was a deployment to GEICO's skills agent platform and a demonstration to his boss at 4:00 p.m. The experiment ran three times across three modernization missions.
The three missions covered uplift, transformation, and reimagination
The rental experiment uplifted an existing system by extending an API. Agents connected information across the domain and found a way to save 8,500 calls per month while giving customers current rental information. The damages experiment transformed the domain that assigns dollar amounts to reported losses by creating agent-callable APIs for a larger workflow. The salvage experiment reimagined the domain covering totaled vehicles. It produced a vetted migration plan, API specifications, and design mock-ups approved by product and technology leaders.
GEICO's knowledge and judgment mattered more than the replaceable vendor components
Prior separates the vendor-provided parts from the GEICO-specific parts. The replaceable side included the harness, model, FleetKit, and code-modernization plugin. GEICO supplied knowledge-base content, software standards, records of past decisions, tool access, usage instructions, development methods, and judgment about who could decide what at each gate. These materials were grouped into repositories so agents and humans could add corrections and new knowledge.
FleetKit turns a mission into research, review, and implementation
FleetKit is a structured collection of markdown rules, protocols, and instructions designed for the Claude harness and architecture transformation. It takes a goal, breaks it down, and dispatches leader agents for research and approach decisions. Results enter a human-in-the-loop queue where people can comment and approve, while agents can add context. Once the approval state is acceptable, builder agents implement the work. The process uses beads to help agents track work inside their own agent spaces and aims to produce a material change in GEICO's repositories or systems.
Worksheets and diagrams gave the war room a shared record of decisions
Each mission used a communication channel with the relevant people invited. At milestones in the nine-step process, the system created worksheets and design documents. The worksheets were HTML artifacts that people could fill in and save. The design documents included architecture visualizations that teams could edit, after which agents reread the updated diagrams and incorporated them into the session. Prior says the artifacts asked useful questions, exposed business and technical edge cases, and documented systems that engineers had not previously taken time to diagram.
The driver still controlled the runtime and paid for orchestration mistakes
The system used cloud agents for code analysis and graphing over legacy domain-specific languages, plus local agents with access to the driver's laptop and enterprise-specific MCPs. The laptop was the runtime, and Prior had to run the process steps. He warns against allowing a Claude session to create cron jobs that launch more sessions, because this can consume the budget immediately. He also reports higher-than-expected prompt caching and says benchmark results showed Sonnet 5.5 wrote code as well as or better than the existing model in that test.
The next improvements are better gates, context search, meeting capture, and rollout
Prior's next priorities include stronger gate enforcement and changes to decision models and model selection. GEICO's context was mostly markdown files, without a sophisticated graph or search system, so he wants a better way to find and inspect it. He also wants to record war room conversations and transcribe them automatically, rather than relying on handwritten notes. For wider deployment, he suggests finding a group that struggles with decisions, controlling costs, driving the system, and adapting it to the organization's own methods.
"So while we've seen this inner loop speed up to use Gene Kim's concept, we find ourselves seemingly stuck within the bounds of this outer loop, which is to say everything else that happens other than the coding and the building."00:56
Who should watch
You are responsible for modernizing a large legacy codebase and need to connect domain knowledge, architecture work, approvals, and implementation.
Your AI coding tools are fast, but product decisions and sign-offs still delay major changes.
You want a concrete pattern for running multiple agents with human approval, internal tools, and organization-specific rules.