Ford Prior, a senior staff engineer in GEICO’s claims technology group, describes three AI-assisted modernization experiments. An 18-minute talk on AI Engineer.

Speed up decisions around the coding

  • Prior says coding accelerated while context gathering, collaboration and sign-offs continued to constrain major changes.
  • His “war room” combines relevant experts and decision makers, one driver, shared communications and fleets of agents.
  • The experiment: spend a day and a $1,000 token budget mapping one claims domain and delivering a concrete change or approved plan.

Three missions produced different outcomes

  • Rentals: extend an existing API; Prior reports identifying a way to avoid about 8,500 calls per month through better rental information.
  • Damages: create agent-callable APIs that expose data needed by a broader workflow.
  • Salvage: produce an approved migration plan, API specifications and mockups; Prior also reports finding an undisclosed production bug.
  • An approved plan and a deployed capability are different milestones; the three missions should be read separately.

Put organizational knowledge beside the model

  • The vendor supplied models, agent-running software and a code-modernization plugin.
  • GEICO supplied historical decisions, standards, tool access, specialized methods and the people authorized to approve each stage.
  • FleetKit, a Markdown-based orchestration rulebook, breaks a goal into work, sends research agents and queues decisions for human approval.
  • Builder agents follow approvals; deterministic checks confirm that required files and other artifacts exist between stages.

Make the collaboration leave useful artifacts

  • A shared mission channel brings in the relevant owner when product or technical details are missing.
  • Interactive worksheets capture answers; editable diagrams and design documents feed back into the agents’ context.
  • Cloud agents analyze legacy code, while local agents use the driver’s connected tools and enterprise systems.

Control spend and strengthen the gates

  • Repeatedly launching an expensive model just to poll progress can consume the budget quickly.
  • Prior reports a lighter model matching or beating a larger model on a local legacy-code comparison once context was prepared.
  • Next steps include stronger approval enforcement, better context retrieval and less lossy capture of meeting decisions.

Caveat: $1,000 is the experiment’s token-budget framing, excluding preparation, participant time and enterprise tooling. FleetKit was described as unreleased, and the talk provides no reproducible evaluation of the three missions.

“Who can decide what and at what gate?” — Ford Prior