Parlant

Parlant lets you define agent behavior as conditional if-then guidelines … add a rule, set its priority, declare which other rules it overrides. The matching engine activates only what applies per turn. No system prompt to maintain at scale, no graph to rewire when a compliance rule changes. Canned responses lock regulated moments to pre-approved wording … no generation, no hallucination. Open source, LLM-agnostic, deployed in banking. #parlant #contextEngineering #conversationalAI #openSource

I’m excited to announce our partnership and integration with Novita AI to bring their inference cloud to Parlant! Novita is one of the top inference providers on Hugging Face, serving 350,000 developers at some of the most competitive pricing in the market. Parlant is now used by thousands of developers and runs in production at major financial institutions, ensuring that conversational customer-facing LLM agents follow behavioral guidelines reliably and consistently, with full traceability of every decision. Together, we're giving teams the ability to build chat agents that are fast, affordable, AND reliably controlled, using open-source code with open-source models. The integration is live now on our development branch, with an official release coming soon. Stay tuned!!

We're thrilled to announce our partnership with Parlant, the open-source conversational AI framework with 18K+ GitHub stars that enterprises are choosing over $10B closed-source competitors for mission-critical chat agents. Novita AI's inference cloud (200+ models, sub-second latency, affordable pricing) is now available on Parlant's development branch and will be officially included in their upcoming major release. Parlant solves the problem that plagues every team building chat agents: reliably controlling what the agent actually says to your customers at scale. Our 350K+ developers can now leverage the most controllable agent framework, while Parlant's community gains access to the most cost-effective, scalable inference.

Parlant 3.2 is out! 🚀 ! The headline: streaming responses. The story is like this... I ran a poll in our Discord server (addressing more than 1k Parlant developers) asking what's their most wanted feature for our next release. The highest score, by far, went to the speed of responses, beating more control (Parlant's pretty damn good at that already), deeper observability, and other options. So in the newly released v3.2, we listened to our awesome community and complied!!! Here's what's new in this release: 🔹 Streaming responses: A new output mode where responses can now arrive token by token instead of as a single block. For voice agents, this means the agent can start speaking before the full response is even ready, which dramatically improves the UX of voice agents built on Parlant. 🔹 Labels: Tag guidelines, journeys, and states with labels that propagate into session metadata automatically. Easily pull up all sessions where an upsell guideline was activated, or where a customer requested escalation to a human. This is our new foundation for real-world analytics on what your agent is actually doing out there. 🔹 Scoped retrievers: Attach data retrievers to specific guidelines or journeys instead of running them globally. For instance, pricing docs only load whenever pricing is discussed. Cleaner context window, fewer wasted calls, and much less of that irritating scope creep agents tend to get. 🔹 Agent personality: You can now easily customize Parlant's unique preamble messages per agent with dynamic, context-aware instructions. 🔹 Track control: Strict control over whether a guideline fires once per conversation or context, or keeps reapplying every time its condition matches. For example, empathy cues and tone adjustments shouldn't be one-and-done. And other useful improvements: explicit field dependencies for canned responses, bulk relationship definitions, and a batch of engine and SDK fixes. Full breakdown on the blog (first comment) 👇

Struggling to build AI agents that actually follow instructions? You spend weeks crafting the perfect system prompt. It works fine in testing. Then real users break it in minutes. The agent ignores rules, hallucinates, and can't handle edge cases. Sound familiar? It's the #1 pain point for AI developers. It's time to stop fighting prompts and start teaching principles. Try 𝗣𝗮𝗿𝗹𝗮𝗻𝘁, the AI agent framework that ensures your LLM follows instructions. With 𝗣𝗮𝗿𝗹𝗮𝗻𝘁, you can: ✅ Define clear customer journeys for your agent. ✅ Craft behavioral guidelines that the agent actually follows. ✅ Attach tools and APIs to specific events. ✅ Use templates to stop hallucinations and keep a consistent style. ✅ Understand exactly why your agent makes each decision. 🔗 Link to repo: github(dot)com/emcie-co/parlant --- ♻️ Found this useful? Share it with another builder. ➕ For daily practical AI and Python posts, follow Banias Baabe.

Most conversational AI systems fail for one simple reason: they’re built like chatbots, not agents. Prompting an LLM isn’t enough when you need: • predictable behavior • multi-step conversations • tool usage • guardrails and consistency In our latest video, We walk through conversational AI agent modeling using Parlant - focusing on how concepts like journeys, guidelines, tools, and state turn raw LLMs into production-ready agents. If you’re building AI assistants, copilots, or support bots, this mental model will save you a lot of pain later. https://lnkd.in/dDGEa6VD

One thing that LLM agents can't do well: Any old-school chatbot can stay on script, while LLM agents tend to go rogue and lead customers into weird conversations. But of course, old chatbots feel robotic, and customers don't want to talk to them. They are reliable, but people don't like them, LLM agents are the opposite. They're fluid and adaptive, but they can say anything. You're literally one hallucination away from a disaster. The guys behind Parlant are doing something really smart with their new version: You can build an agent with the best of both worlds. The agent can dynamically switch between an LLM agent and strict mode based on what's happening in the conversation. Risk isn't uniform across a conversation: 1. When a customer asks a casual product question, Parlant engages the LLM to generate a fluid and helpful answer. 2. When a customer asks for a refund, Parlant engages strict mode to only return approved, contextually-driven response templates. You control the agent's "composition mode" based on natural-language observations about the current state of the conversation. This is a really cool idea. It should significantly improve the current state of the art in chatbots. You can check it out here: parlant.io The attached diagram shows how the dynamic composition mode works.

Shout out to my friends at Emcie

🚀 Launching Parlant 3.1 [TL;DR: Speed, Accuracy, & Control] Parlant has seen incredible adoption in the past few months. It's live in production with a few large financial institutions, and it also helps many small-to-medium businesses in use cases from healthcare, through legal, all the way to education and - get this - live troubleshooting of wind turbines. There are more use cases that our open-source Conversational AI framework seems to excel at than we first had in mind. Today - I'm excited to launch Parlant 3.1! Since releasing 3.0 a few months back, we've listened to our community and collected an enormous amount of feedback and feature requests. The additional feature list in this version is extraordinary. In terms of the experience, expect improvements in speed, accuracy, and — the real keyword of this release — CONTROL. We've added multiple new control mechanisms in this new version that leverage and really show the power of Parlant's granular behavior control philosophy. Check out the expanded feature overview in the first comment!

The hardest part about building agents isn't the reasoning, it's the reliability. You eventually hit a wall where adding one new instruction to your system prompt makes the model forget an old one. It’s a game of whack-a-mole. I’ve been digging through the Parlant repo, and they seem to have solved this by flipping the architecture entirely. Instead of dumping a 2,000-word essay into the context window, they built an engine that sits between the user and the LLM. They call it the Guideline Matcher. The engine scans the conversation state first, finds the specific rules (guidelines) that apply to that exact moment, and only then instructs the LLM. It completely solves the context pollution problem. => If a user is asking for a refund, the LLM doesn’t need to know the rules for "Technical Support." => It creates a strict separation of concerns: The LLM handles the conversation, but the Engine handles the logic. It feels less like prompt engineering and more like writing guarded code. It treats agent behaviors as strict logic, not fuzzy suggestions. Exactly what the stack was missing. Here's the repo: https://lnkd.in/gGNMRGKR [Do support them with a ⭐️] ✔️ Follow Sandhya for more AI insights.