An agent that reads the raw mess — carrier emails, webhooks, a photo of a damaged label — resolves the shipment dynamically, recovers from failures, and verifies the fix against real state.
Every run starts from a raw artifact — an autogenerated carrier email, a half-documented JSON webhook, or a photograph of a damaged label. The agent extracts the shipment and the exception under ambiguity, with vision when the input is an image.
State changes are idempotent. Transient failures retry with exponential backoff; permanent failures fail fast and escalate. The act-versus-escalate decision is driven by a real failure class in code — not a hopeful prompt.
After any change the agent re-reads authoritative state and returns a field-level before/after diff. You see exactly what moved and a verified resolved — or a deliberate escalation to a human. No unverified claims.
A hand-rolled agent loop. One tool call per turn, provider failover underneath, a structured trace on top. Watch it resolve three different exceptions along three different paths.
Pick a scenario, run the agent, and watch it stream every step — then read the verified diff. Nothing here is faked.