Autonomous Operations Agents

Context

Two operations agents, both built on the Claude Agent SDK, running different sides of the business:

• Feral Golf Co. - an always-on agent covering customer support, sales, builder/manufacturing-partner management, and marketing, letting the brand run these functions without adding headcount as it scales.
• California Seafood - a delivery operation where drivers and customers text in change requests, substitution responses, order confirmations, and availability questions, all of which a human operations coordinator was triaging manually against a delivery-window cutoff schedule and the company’s ordering system of record (Cut & Dry).

What I built / did

Feral Golf Co.: a single agent handling customer support inquiries, sales conversations, coordination with manufacturing/build partners, and marketing tasks - the operational layer that lets Feral Golf run lean.

California Seafood: an autonomous operations agent that:

• Classifies inbound SMS messages into five scenario types - customer change requests, responses to agent-proposed substitutions, order confirmations, ambiguous/cancellation requests, and availability inquiries - and routes each one to the correct handler.
• Enforces delivery-window cutoffs in Pacific time, automatically acknowledging requests inside the window and escalating anything outside it to a human.
• Runs the substitution flow end-to-end: checks live availability, proposes substitutions, and auto-confirms them if the customer doesn’t respond before the window closes.
• Respects a strict data-integrity rule: Cut & Dry (the source of truth) is always written first; the driver-facing Google Sheet is only mirrored after that write succeeds - never the other way around.
• Designed the system so ambiguous or cancellation-adjacent messages are always escalated to a human rather than guessed at.