Zendesk for StartupsUp to 2 years free for startups + AI Agents & Copilot
Promoted
Launched this week
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%.
Would love to look under the hood to see how the agents are learning. Part of me thinks it's just a loop with markdown files haha, well that's at least how I've got my agents learning.
Hey Yi, congrats on the launch! I totally get how frustrating manual log tracing can be when you are trying to fix those pesky failing prompts in production.
I am curious, how does Reflexio tackle prompt drift or regressions on edge cases that are already working when it goes through its auto optimization process? I can nott wait to see how this scales up!
Reflexio
@samar_cheema1 Reflexio treats the agent as a black box and continuously observes how the agent behaves with the learning injected. If there is any regression for learned playbooks, it will automatically update those learned playbooks to reflect the new agent prompt or user behavior distribution.
We also have an offline reinforcement learning loop that will proactively detect those cases.
Yi, congrats on the launch.
I know closing the online feedback loop for agents in production can be a real headache manually sifting through traces and adjusting prompts can get tedious pretty quickly.Just a quick question: when Reflexio optimize behavior is it dynamically tweaking the prompt or few shot context or is it fine tuning the models behind the scenes?
That 57% reduction in tokens with fewer failures is impressive! Wishing you all the best with the launch
Reflexio
@janet_gardon We treat the agent as a black box. So agent will search for additional context from Reflexio to steer the agent responses. Those learnings from Reflexio help the agent to avoid making the same mistakes it made across all production traffic
Clipto
Congrats! Does Reflexio learn per-user, per-agent, or globally? Curious how you separate a correction that's specific to one person from one that applies to everyone.
Reflexio
@henrykang thanks for the question, reflexio learns user profiling information and user level playbook at first. When there is a pattern detected across users, we will find the agent-level and global-level playbooks that can help any user, including new users. Therefore, Reflexio can help the agent quality cold start problem as well!
Congrats on the launch. The 30% failure drop and 60% token savings numbers are strong, how did you actually measure those? Was it an A/B between agents with and without Reflexio on the same task set, or before/after on the same production traffic?
Reflexio
@wwenioaguiar Yes, those are measured across our customers through A/B tests, comparing the agent response with Reflexio and without on the same production traffic.
Dial
interesting approach - the generalized-rule part is what gives me pause though. if one user's "correction" is actually bad advice (they misunderstood the task, or just pushed the agent toward a wrong answer confidently), and that gets rolled up into a rule applied to every user, how do you catch that before it spreads? is there some confidence threshold or does every generalized rule need a human to sign off before it goes from "one user's pattern" to "everyone's default behavior"?
Reflexio
@galdayan You are talking about exactly the core value we are providing! Reflexio discover and judge if this pattern should be generalized to every user.
There are two gates on there:
1. We check if there is a certain user group that shares the same feedback consistently. If so, that is unlikely to be a one-off or bad advice from a single case, but more of a generalized pattern for that user group.
2. Those agent feedbacks can be reviewed by a developer, and only approved agent feedbacks will become everyone's default (if decided to be on the conservative side)
Dial
@yilu the human sign-off gate is the key piece - glad it's not fully automatic. makes sense now, thanks for walking through it