Viktor.comAn AI coworker that actually does the work
Promoted
Launching today
Reflexio
Behavioral learning that makes AI agents better over time
96 followers
Behavioral learning that makes AI agents better over time
96 followers
Reflexio makes AI agents better with every interaction. When users correct an agent, a path fails, or something works particularly well, Reflexio turns that experience into behavior the agent can reuse next time. Instead of leaving valuable lessons buried in logs, your agent continuously learns what to repeat and what to avoid — with every learning visible, testable, and reversible. Reduce task failure rate by more than 30%, while saving tokens by more than 60%.
Reflexio
Hey Product Hunt 👋
I'm Yi, co-founder of Reflexio. Before starting Reflexio, I was tech lead in Meta and adjunct professor at University of Washington teaching ML and business applications.
Today we're launching Reflexio: a learning platform that makes your AI agents fail less and burn fewer tokens, by learning from what actually happens in production.
Here's what got us started: people use AI agents every day, but agents never actually get better with use. Even with memory, an agent that failed a task yesterday will fail the same way today, across different users — because nothing connects what happened in production back to how the agent behaves next time. The online learning loop just isn't there.
We learned firsthand that closing that loop manually — reading traces, spotting failures, rewriting prompts — is a painful, never-ending job.
Reflexio autonomously observes your agent's live traces, learns from successes, failures, and user corrections, and continuously optimizes behavior. No manual tuning.
The results? In our case studies, agents with Reflexio:
🎯 Cut task failure rate by 36%
💸 Reduced token usage by 57%
📈 Improved response quality in 47% of interactions, with negligible regressions
Try it today: sign up free at reflexio.ai and get 30 days of Pro on us.
@yilu It's a real challenge when agents keep making the same mistakes across various sessions and users. Once you start scaling, manual trace analysis can quickly become overwhelming. Just a quick question about the token reduction—are you mainly achieving that through more streamlined and optimized prompts over time, or are you also cutting out any unnecessary context?
Reflexio
@liam_ward2 Awesome question. You're getting to the core of this product.
Because the agent has worked on the similar task before, Reflexio will help the agent find out what is more optimized to follow next time. This means 75% fewer internal model steps and roughly translated to 57% token reduction because less detour agent needs to make next time. We validate our impact with real customer case study https://www.reflexio.ai/blog/creao-repeatable-agent-execution and GDPval dataset
nice，can I export or delete all learnings if I decide to move off the platform
Reflexio
@peggykk absolutely, all of those user profiles and learnings are in your control and can be exported if you want to move off from it. Note that, even though learnings are generated, we have offline reinforcement learning pipeline that continuously optimize learned signals. So as there are more traffic, existing learnings will be updated as well to find even better playbook to follow
Reflexio
@peggykk Yes, you can delete a user account along with all associated learnings and metadata from Reflexio. We also support exporting all learnings upon request.
CapCut AI Suite
Saw Python, REST, and CLI, do you have any plans for a TypeScript SDK?
Reflexio
@lavana_cricko good question, yes, typescript is on our roadmap! stay tune to it. in the meanwhile, REST api endpoint is probably your best bet. let us know how it works for you and reduce your agent's failure rate!
NexaSDK for Mobile
I think I got the concept.
How do you measure "negligible regressions"? Is there an eval harness that runs before a learning gets applied?
HyNote AI
Can I review and approve learnings before they go live, or is it completely autonomous?
Reflexio
@ristan_nakko great question. You can definitely first publish user interactions and see all the learnings from the dashboard, or use the API to check them. If you don't like them, you can edit them, delete them, or even adjust your learning objective and re-generated those learnings based on your customerizations!
How much control do developers get over which lessons the agent keeps?
Reflexio
@jackthompson68 We believe that everyone's agent is a little bit different, so customization is our first-class citizen. Developers can control:
what reflexio should focus on to learn
Context window that the learner should learn from
Frequency of the learnings
How aggressively to aggregate user-level learning into agent-level, which generalize to all users, include new ones
Developer can also see, edit, and delete all the learnings before adding them to the system.
Reflexio
The journey of building Reflexio started from a painful lesson Yi and I learned firsthand.
At our previous company, we worked on the personalization service and memory infrastructure powering AI agents at very large scale, and saw how hard it is to make agents actually learn from experience. It wasn’t just about storing user facts. Teams spent huge amounts of time reviewing production conversations: where agents failed, where users corrected them, which tools were called incorrectly, and how those lessons could be turned into better behavior through metrics, evaluations, and experiments.
That kind of learning infrastructure is powerful, but it takes serious engineering investment — the sort only a handful of companies can afford. Most agent builders and startups don’t have a dedicated platform team of that size behind them.
That became our “aha” moment: as AI agents become more useful, every team will need a way for agents to learn continuously from real interactions, changing environments, and user feedback.
Earlier this year, we left to build Reflexio: a learning platform for AI agents. We prototyped it locally, offered it as a cloud service, integrated it with coding agents like Claude Code and Codex, and worked closely with design partners to refine the product through real usage.
Today, we’re excited to launch Reflexio on Product Hunt. It’s already being tested with design partners and enterprise customers, but this is just the beginning. If you’re building AI agents and believe they should get better every time they interact with the world, we’d love for you to try Reflexio, share feedback, and join us on the journey.
Reflexio
@guangyu_yang1 yeah, at Meta, I literally have to assign a number of engineers to do production online learning, and that is an endless job for all the AI features we have deployed. We are automating the online learning for AI agent for everyone!