Perplexity says it has trained AI agents to learn from real user sessions, including mistakes, and reports that a resulting checkpoint reduced tool-call failures by 21.2%. The work offers a practical glimpse of how agent developers might turn user complaints and tool errors into better decisions, rather than simply teaching models to imitate successful outcomes.
Perplexity Watch analysis
What happened
In a post published on 12 November, Perplexity researchers describe a training pipeline combining rejection-sampling fine-tuning with hint-guided self-distillation. The system uses signals such as user complaints and tool errors to identify avoidable mistakes, then generates corrective hints for training.
The company says its GLM-5.2 checkpoint achieved a 21.2% reduction in tool-call failures. It also reports improved performance in offline benchmarks and online A/B tests. The post does not provide the evaluation details or figures behind those broader results in the supplied account.
Why it matters
Agents often fail in the middle of a task: a tool call goes wrong, or the model makes a poor decision before reaching its final answer. Perplexity’s method targets those intermediate steps, using evidence from real interactions to identify where the process went off course.
If the approach generalises, it could give developers a way to improve agents from both successful and unsuccessful sessions. That is useful beyond one checkpoint: tool reliability is one of the less glamorous problems that determines whether an agent can be trusted with actual work.
Our read
A reported 21.2% drop in tool-call failures is a concrete result, and learning from failed attempts is a sensible direction. But the number needs its evaluation context before readers can judge how large or broadly useful the improvement is. For now, this is a promising account of Perplexity’s method, not a universal recipe for fixing agents.
What to watch
- The evaluation setup and baseline behind the reported 21.2% reduction.
- Whether Perplexity publishes results across more tasks and models.
- Whether the method improves reliability in deployed agents, not only in its reported evaluations.
Discussion spark: Should agent developers use real user complaints and tool errors to guide training, or does that risk making systems optimise for the wrong signals?
Sources and evidence
- Learning from Real-World Experience (12 November 2026, 01:29 UTC)
Independent WittyWires tracker for public updates about Perplexity. Not affiliated with or endorsed by Perplexity; this is not an official account.