I've been using them for months and felt in love on first sight. Great product by UX/UI and also very useful - nice to see them their innovation pace is so fast too! Looking forward to see them in great places!
This looks like a really smart way to build AI agents without all the usual headaches! I love that it can refine prompts automatically, does that mean it learns from past mistakes and improves over time?
@heycesrThis looks really cool, César! The self-improving part is especially interesting. AI agents that learn and refine themselves over time instead of staying static could be really amazing.
How do you prevent agents from "over-correcting" or drifting too far from their original goal? Does the system track changes to ensure the improvements actually lead to better results?
Replies
Latitude
Hello Product Hunt!
We're happy to come back to PH to introduce Latitude AI Agents, the end-to-end platform to design, evaluate, and refine your agents.
Key Features:
- Autonomous Agentic Runtime: Craft prompts that run in a loop until the agent achieves its goal, fully integrated with your existing tools and data.
- Multi-Agent Orchestration: Break down complex tasks into smaller agents and easily manage their contexts.
- Self-Improving Prompts: Use other LLMs to evaluate agent performance and automatically refine the agent's instructions based on the results.
- Easy Integration via SDK or API: Integrate with agents into your codebase using our SDKs for Python and TypeScript.
- Model Context Protocol Ready: Connect with many platforms offering tools and resources for agents, or create your own custom MCP server.
We'd love to hear your thoughts. Are you building an agent in 2025?
Looking forward to your feedback!
Awesome stuff!
Pullpo.io
Hey César! This is cool! How does agent performance evaluation work? I imagine it can sometimes be really hard to do, even for a human.
Latitude
@marco_patino Thanks Marco! We have a range of evaluators available:
LLM-as-judge: an LLM analyzes the instructions and list of messages your agent produces
Human in the loop: a human reviews agent generations and scores them manually
Code evals: you can push evaluation results directly from your backend
It really depends on the use case, but we've seen improvements of up to 30% using our automatic prompt refiner.
Pullpo.io
@heycesr nicee 👌
The "self-improving" part is interesting, could be useful for creating agents that get better over time instead of staying static.
Great idea. How does the self improvement part work?
Latitude
@ed_preble We use a technique called Semantic Backpropagation:
You can evaluate any conversations generated by the agent automatically using LLM-as-judge
We use the results of those evaluations to suggest changes in your prompt automatically
If you want to learn more, I highly recommend this paper: https://arxiv.org/pdf/2412.03624
Thinkbuddy AI
I've been using them for months and felt in love on first sight. Great product by UX/UI and also very useful - nice to see them their innovation pace is so fast too! Looking forward to see them in great places!
Fable Wizard
This looks like a really smart way to build AI agents without all the usual headaches! I love that it can refine prompts automatically, does that mean it learns from past mistakes and improves over time?
All the best with the launch! @Latitude @heycesr
By this product Ai agents building app will be accessible to a wider audience and performance of such agents can be evaluated in realistic scenarios.
@heycesrThis looks really cool, César! The self-improving part is especially interesting. AI agents that learn and refine themselves over time instead of staying static could be really amazing.
How do you prevent agents from "over-correcting" or drifting too far from their original goal? Does the system track changes to ensure the improvements actually lead to better results?
Congratulation for the launch!