Parlant

Parlant is an open-source framework specifically designed for the messy reality of customer-facing chat agents: shifting business rules, multi-turn conversations, compliance constraints, and the need for humans to steer the bot without rewriting prompts. Instead of one giant system prompt, you give Parlant a set of declarative guidelines (e.g. 'when the user asks about refunds and the order is older than 30 days, escalate to a human') and the framework decides at runtime which guidelines apply to the current turn, how to combine them, and whether to call a tool, ask a clarifying question, or hand off. This gives product and support teams a much more auditable and editable way to shape agent behavior than monolithic prompt engineering. Parlant ships a Python SDK, a session manager, tool definitions, evaluation harnesses, and an inspector UI to show why the agent did what it did. The project has crossed 18k GitHub stars (3.3 released in mid-2026 with features for 'managing complexity at scale') and is used in production at Chase's ML org and at Slice Bank, among others. It is a strong fit for support, sales-qualification, and onboarding bots where 'mostly works' isn't good enough and you need to debug and constrain behavior at scale.