Just Now Possible Automating the Full Customer Support Iceberg: How Gradient Labs Built a Multi-Agent Platform
36 snips
Dec 18, 2025 Jack Taylor, a product engineer at Gradient Labs, and Ibrahim Faruqi, an AI engineer, dive into the complex world of customer support automation. They discuss the iceberg metaphor of support, emphasizing how much more lies beneath the surface. Unique agent coordination for inbound, back office, and outbound tasks is explored, along with natural language procedures allowing experts to train agents seamlessly. The challenges of defining success in outbound interactions and ensuring regulatory compliance through innovative guardrails add intrigue to their enlightening conversation.
AI Snips
Chapters
Transcript
Episode notes
Let Experts Write Procedures In Plain Language
- Let subject matter experts write procedures in natural language so agents follow real company workflows.
- Avoid forcing non-technical experts to translate processes into code to reduce knowledge loss and speed deployment.
Outbound Needs Its Own Completion Logic
- Outbound agents must define 'done' because customers don't always signal completion.
- Determining completion requires different logic than inbound flows where customers end the interaction.
State Machine Turns Coordinate Long Conversations
- They use a state machine orchestrator that triggers long-running turns and preserves conversation state.
- Turns invoke modular skills to decide next actions across async, multi-day exchanges.