TripFix — autonomous flight-claim co-pilot

Early on the obvious move was “one model, one prompt, one pass.” That breaks the moment the model gets a passport photo upside down or sees an email it’s never seen before. I split the work into specialised agents— one reads documents, one drafts the rebuttal, one decides when to escalate to a human, one judges the output.

The agents talk through a typed event bus, not a tangled chain. Each one is testable in isolation, and any one of them can be swapped (or quietly rolled back) without touching the others.

• Document agent reads, classifies, extracts structured data
• Drafting agent composes the airline rebuttal with citations
• Routing agent decides what a human should see
• Judge agent scores the output before it’s sent

Letters that quote the wrong booking reference are worse than no letter at all — they break trust instantly. So the drafting agent cannot output a claim without a citation handlethat maps back to a span in a real document.

If the agent wants to say “the airline confirmed the cancellation at 14:32,” it must point to the line in the email where that confirmation lives. The UI renders those citations as clickable chips; the eval harness checks every one.

Every prompt change goes through a five-evaluator harness that scores hallucination, citation grounding, tone, completeness, and refusal handling. If the new prompt regresses on any dimension, the deploy is blocked.

This is the only reason one engineer can iterate on prompts daily at production scale without quietly breaking the product.

Agentic systems live or die by how fast a human can audit a bad run. Every agent at TripFix writes to a timeline that surfaces its prompt, its thinking, every tool it called, and the answer. If an ops human can’t debug a run in 60 seconds, we don’t call it production-ready.

No single model is the best at everything — or the cheapest. The router picks the right LLM for each sub-task based on latency, cost, and historical eval scores. Frontier-class for the rebuttal draft; smaller models for classification and routing.

Result: production cost per claim stayed flat as the system did more work, and any single-vendor outage stops being existential.