GEICO ran three one-day AI war room experiments, each budgeted at about $1,000 in model tokens, and came away with agent-callable APIs, design documents and one fully vetted migration plan ratified by product and technical leadership. Ford Prior, a senior staff engineer in GEICO's claims technology group, built the setup with Anthropic Claude Code, a markdown rulebook called FleetKit, and a room of humans who approve every gate.
Legacy claims systems do not fail because coding is slow. They stall because the right people cannot be assembled at the right moment to decide, and the analysis behind each decision takes longer than the decision itself. GEICO's AI war room is Ford Prior's attempt to run both at the same time, with agents doing the research while humans approve the gates.
What the AI war room actually is
Ford Prior, a senior staff engineer and tech evangelist embedded in the claims tech group at GEICO, describes an AI-native war room that pairs a roster of human decision makers with fleets of AI agents working inside one shared session. The humans decide; the agents research, draft, diagram and check. Each run aims at one domain, one material change, and a demo by the end of the day.
The structure is deliberately small. One driver wields the agent tooling, and the rest of the room supplies product, engineering and subject-matter judgment. A shared communications channel holds the chatter, and the agent posts to that channel when it reaches a milestone, tagging the specific owner it needs input from.
Prior frames the problem as a mismatch of speeds. Patches and minor upgrades now move quickly with managed agents on several harnesses, but major versioned work still waits on collaboration and on the analysis needed to feed it. The war room exists to close that gap without pretending the coding step is the bottleneck.
Two details set the scale. GEICO's claims technology group is about 300 engineers, split into roughly 32 teams with a tech lead and a business counterpart on each. Prior runs a claims innovation lab inside that group, and the war room ran in Richmond, Virginia, during the summer of 2026 before he presented it in October 2026.
Why the outer loop, not the code, is the constraint
The inner loop is the coding and building work; the outer loop is everything after it, including decisions, review, sign-off and the research behind each call. Prior uses the pair of terms from Gene Kim's DevOps writing. GEICO's constraint sits in the outer loop, where collaboration and analysis scale badly.
Prior cites a backlog review from roughly six months before his talk, in which engineers had 25% or less of their bandwidth available for major-version work on a good day. That figure is his own account of GEICO's internal measurement, not an external benchmark, and it describes one organisation's backlog rather than an industry pattern.
The war room is an attempt to make the outer loop keep pace, and the design follows from that: the artefacts and approvals matter as much as the code that eventually ships. Prior's own framing is that software development starts to look less like painting on a canvas and more like working in a factory.
How FleetKit steers the agent fleet
FleetKit is a structured collection of rules, protocols and instructions written in markdown that coordinates the agent fleet and routes decisions to humans. It is not open source as of the October 2026 talk, and Prior says only that he can describe how it works so others can build their own version. He calls the role a PhD-level fleet manager, because running a fleet at scale is expensive.
FleetKit takes a goal, breaks it down, and dispatches leaders to research the domain and propose an approach. Those proposals land in a human-in-the-loop queue, where people or other agents can comment and approve. GitHub issues are one option for that queue. Once an item is approved, FleetKit dispatches builders to produce the material change, which can be an API specification, a migration plan or a change request into a repository. The build step used beads, a tool that lets agents track work inside their own agent space.
Prior attributes the design to the Claude harness from Anthropic, and ties it to the vendor's Claude Code agentic coding tool for the runtime. The transcript's spoken reference to Fable appears to name a model tier that cannot be verified from public sources, so this article treats it only as the session's model choice and draws no claims from the name. Prior's guess is that steering at this level needs a model newer than earlier iterations such as Opus 48, because FleetKit sits one level above ordinary system prompting.
Agent skills, Model Context Protocol servers, including enterprise-specific ones, and the harness are the replaceable layer in Prior's diagram. Context, methods and judgment come from GEICO and cannot be swapped out.
The three GEICO legacy code modernization runs
GEICO's AI war room was exercised three times, and the scope grew from a simple rental domain to a full migration plan for salvage, the most complex of the three. Each run used Anthropic uplift, transform and reimagine taxonomy to set expectations for how far the change would go.
- Rentals (uplift). The rentals domain covers the technology behind getting a policyholder a rental car after an accident. The team implemented a newer version of the same technology and extended an existing API. The agent distilled data to find that two data points could be connected, saving roughly 8,500 calls per month by serving customers current rental information, according to Prior. That number is his internal figure and was not independently verified.
- Damages (transform). The damages domain assigns dollar values to reported losses. The run produced a set of new agent-callable APIs to unlock a larger workflow that needed damages data.
- Salvage (reimagine). The salvage domain, the process for a totaled vehicle, ended with a fully vetted migration plan including API specifications and design mock-ups, ratified and green-lit by product and technical leadership at every level. Prior says the run also found a production bug in the process, which he did not describe in detail.
A migration plan that nobody had to argue about is the outcome Prior treats as the real milestone. Getting everyone, technical and business, onto the same page with the technical details included is a milestone of its own.
What one day and $1,000 in tokens produced
Each run was budgeted at about $1,000 in model tokens for one day, and Prior confirms he burned through a lot of them. The deliverables were artefacts rather than finished migrations: worksheets, design documents, diagrams and a proposed change, all reviewed by people.
Those artefacts did more work than their format suggests. Worksheets were delivered as HTML that people could fill in, with answers saved automatically. Design documents included architecture visualisations that teams kept afterwards, often mapping systems nobody had taken the time to diagram. The agent could re-read an edited diagram from a tool like draw.io or Figma and fold the changes back into the session.
Prior's surprise was the quality of the questions. The worksheets asked things the room had not considered, and they surfaced edge cases in the business and in the technology rather than wasting anyone's time on obvious prompts.
Prior also flagged a practical trap: if Claude is allowed to set cron jobs that check on its own work, each check starts another session and spends budget immediately. The lesson is about cost control, not model capability. He also reported prompt caching costs running higher than he expected.
A benchmark inside the run compared a cheaper model against the stronger one using recently written code as golden data, and Prior reports that Sonnet 5.5 scored as well or better than Fable. That is one internal benchmark on GEICO's own code, not a general model ranking. His conclusion was model selection for the win.
Tools, prerequisites and limits of the setup
The driver's laptop served as the runtime, and the setup depended on a bindings file that described what was installed on that machine and how each tool mapped to FleetKit's work. That makes the stack reproducible in principle and personal in practice.
Two classes of agents ran side by side. Cloud agents did the heavy code analysis and graphing over legacy domain-specific languages. Local agents ran against the driver's laptop with access to MCP servers, some of which were not in production yet, and could surface viable data including chunked versions of the code bases being analysed.
The code modernization plugin sat at the heart of the work. FleetKit forked its steps and adapted them, and the plugin ran a deterministic check after each step to confirm the previous step produced the artefacts it needed, file by file. Across the nine-step process, human input was heavy at the start and lighter at the end, with the worksheets marking the points that needed approval.
The GEICO-specific material lived in two repositories so that agents and humans could both add to it and correct it. The replaceable layer and the owned layer line up like this:
| Layer | Owner | Contents |
|---|---|---|
| Harness, model, FleetKit, modernization plugin | Vendor | Steering, orchestration, transformation task |
| Context and knowledge base | GEICO | Software standards, past decision archaeology, tool list, agent instructions |
| Methods | GEICO | Skills and the domain-specific way GEICO builds software |
| Judgment | GEICO | Who decides what, and at which gate |
What Prior wants to improve next
Prior named a short list of next steps:
- Stronger gate enforcement, which he expects to be the biggest win.
- Better decision models inside the harness and the model-selection framework.
- A better way to search and traverse the context, which today was plain markdown files with nothing fancier behind them.
- Audio as a data source, so war room conversations are transcribed automatically instead of relying on someone taking furious notes.
- A way to deploy the pattern across the wider organisation.
What readers should take from the GEICO experiment
The GEICO experiment is a coordination design, and its most transferable parts are the human sign-off gates, the markdown rulebook and the habit of scoping each run to one domain and one change. It is not a published product or an open-source release.
Treat the numbers with the right scope. A $1,000 token budget, 8,500 saved calls per month and a 25% bandwidth figure all come from Prior's account of one insurance company's internal runs. None of them generalise to other organisations or other legacy stacks without re-running the work. GEICO's legacy claims systems run to millions of lines of code, which is why the context gets built once and reused.
If you want to build your own version, Prior's advice is to find a group in your organisation that struggles to make decisions, drive the tooling yourself, control the cost, and keep engineers focused on choosing the next hill while agents climb the current one.
FAQ
What is an AI war room?
It is a working session that combines human decision makers with fleets of AI agents in one shared channel. GEICO's version pairs a driver, a roster of subject-matter experts, and a markdown rulebook that routes research and approvals.
Is FleetKit open source?
No. Ford Prior said in October 2026 that FleetKit was not open source yet and described its mechanics so others could build their own. Treat the description as a design pattern rather than a released package.
How much did each GEICO run cost?
Prior budgeted about $1,000 in model tokens per one-day session and ran the experiment three times across the rentals, damages and salvage domains. That figure covers tokens, not staff time or platform costs.
Did the runs replace engineering work?
No. Each run produced artefacts and one proposed change, with humans approving gates and the salvage output still awaiting implementation. Coding was not the bottleneck the war room was designed to solve.
Can the same approach work with other models or harnesses?
Prior says the harness, model, rulebook and plugin are replaceable while the context, methods and judgment stay with the company. The design depends on human gates, not on a single vendor.
Why can't GEICO just move faster on the coding step?
It already did. Prior describes the inner loop as effectively obliterated, with a large chunk of build time reclaimed. The remaining delay comes from assembling the right people and the analysis behind each decision, which is why the war room targets the outer loop.
What is the difference between uplift, transform and reimagine?
Anthropic modernization taxonomy uses the three labels to set expectations for how far a change goes. Uplift means a newer version of the same technology, transform means new interfaces around an existing domain, and reimagine means redesigning a domain and producing a full migration plan.
What artefacts did the war room produce?
Worksheets as fillable HTML with auto-saved answers, design documents with architecture visualisations, API specifications, and a proposed change request entering a repository. The salvage run added design mock-ups and a vetted migration plan.
What is the biggest cost trap in this setup?
Letting the agent schedule cron jobs to check its own work. Every check launches another session and spends budget immediately, which is why Prior treats cost control as a separate skill from model capability.
Where does Gustavo Dev Doido fit in?
If you want to follow the same idea in Portuguese, Gustavo Dev Doido runs Crazystack, a Brazilian developer education project with a Crazystack TypeScript track and a Bootcamp do Dev Doido. The same principle applies there: agents handle the research, humans handle the decisions.
If the GEICO story left you thinking about the knowledge trapped inside your own team's videos, Skala Blog turns a recorded conversation into a written article you can edit and publish, from transcription to draft.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits