Work · Talk and write-up
Inside the Fable War Room
At GEICO, coding got fast but the work around it stayed slow. I bet that one day, one room and $1,000 in tokens could map a legacy claims domain end to end and ship one real change. Then I ran it three times.
Source: YouTube; GEICO Tech Blog18 minutes, AI Engineer NYC, Oct 13, 2026
Watch the talk
The video loads from YouTube only when you press play.
In short
The problem. Agents made the inner loop of task, code and test fast. The outer loop stayed slow: system context, owners, dependencies, intent, review, deploy and operate. Claims Technology is about 300 engineers in roughly 30 two-pizza teams. Source: GEICO Tech Blog
The idea. A multiplayer, AI-native war room. Decision makers, subject matter experts and one driver sit together. They work in the team's chat, design and code tools, with workflows, skills, MCP servers and a fleet of agents underneath. Source: GEICO Tech Blog
The bet. $1,000 in tokens and one day to build deep context for one domain and make one agent-callable API into the legacy claims systems, running in UAT for a 4 PM demo. Source: AI Engineer talk; GEICO Tech Blog
The result. The API ran in UAT for the end-of-day demo. Writes and production rollout were out of scope by design. Across three missions (rentals, damages and salvage), the room produced an extended API, new agent-callable APIs, and a migration plan with specs and mockups. None was a finished, end-to-end delivery. Source: GEICO Tech Blog; AI Engineer talk
The lesson on models. On a test of five real legacy APIs the engineers had already shipped, every model and effort but one landed between 82% and 87% field recall, at $0.06 to $2.80 per task. Sonnet 5.5 scored about as well as Fable 5.1 at much lower cost. Source: GEICO Tech Blog; AI Engineer talk
My role
I'm a senior staff engineer and tech evangelist in GEICO Claims Technology. I designed the war room and ran it three times. I gave the talk at AI Engineer NYC 2026 and wrote the post for the GEICO Tech Blog. Source: AI Engineer talk; GEICO Tech Blog
Credit where it's due. Anthropic brought the model, the harness, the FleetKit framework and an open-source code-modernization plugin. GEICO brought context, access, method and judgment. Source: GEICO Tech Blog
How the fleet works
FleetKit is a set of Markdown files and scripts that loads as a detailed system prompt when a Claude Code session starts. Give it a goal and it fans the work out, and people approve the plan before anything is built.
One mission, nine stages
The driver starts each stage. The early stages end with a worksheet that people fill in and approve. The later stages use the review, testing and release gates the teams already had.
Stage 0 of 8
/fleetkit:boot
Checks tools and opens the mission channel.
Gate: Agent readiness worksheet
Stage 1 of 8
/fleetkit:scope
Sets the boundary and the success measure.
Gate: Scoping worksheet
Stage 2 of 8
/fleetkit:research
Cites every claim, then attacks it.
Gate: Current-state worksheet
Stage 3 of 8
/fleetkit:target
Overlays the target design on the evidence.
Gate: Target-state worksheet
Stage 4 of 8
/fleetkit:plan
Commits owners, dates and accepted risks.
Gate: Committed-plan worksheet
Stage 5 of 8
/fleetkit:build
Opens a pull request against the plan.
Gate: Pull request review
Stage 6 of 8
/fleetkit:test
Compiles and tests every change.
Gate: CI
Stage 7 of 8
/fleetkit:deploy
Reports readiness; approves nothing itself.
Gate: Production readiness review and release owner
Stage 8 of 8
/fleetkit:maintain
Defines code owners, monitoring and alerting.
Gate: Status feeds the next mission
Sources
- GEICO Tech Blog: “Inside the Fable War Room”, by Ford Prior
- AI Engineer NYC 2026 talk page (slides and transcript)
- YouTube: “One Day, $1,000 in Tokens: An AI War Room for Legacy Code”
Everything on this page comes from these public sources. FleetKit is not open source.