Kognar Labs · Research
Kognar Labs publishes the work behind the Kognar Platform: how autonomous agents reason, retrieve, coordinate and stay reliable in production. Every paper comes with code, data and the honest results.
What we study
Our research agenda is shaped by what breaks when agents meet real work. Each area feeds directly into the platform our customers run every day.
How agents plan over long horizons, recover from mistakes and know when to stop and ask.
Grounding models in private, changing knowledge without losing precision or leaking context.
Measuring agent behaviour the way operators experience it: regressions, drift and silent failures.
Interfaces and protocols that let people supervise, correct and trust autonomous systems.
Publications
We introduce a calibration objective that lets agents trade task progress for clarification, and show a 41% reduction in unrecoverable actions across 1,200 real operations workflows, with no loss in completion rate.
| Method | Unrecoverable actions per 1,000 tasks |
|---|---|
| Baseline | 164 |
| ReAct | 138 |
| Reflexion | 121 |
| Stop-and-Ask (proposed method) | 71 |
Showing 6 papers
How we work
Every result we publish has already survived contact with customer workloads inside the Kognar Platform. We release the code, the datasets and the negative results, because research you can’t reproduce isn’t research.
Lab-only wins don’t count. We measure against the workflows our customers actually run.
Negative results and ablations ship with every paper. They’re usually the useful part.
Each release includes a runnable repository, the evaluation harness and the raw data.
We study supervision as a first-class problem, not as an afterthought to autonomy.
Research digest
One email per release. No marketing, no digests of digests. Just the paper, the code and a two-paragraph summary.
Or follow the RSS feed · arXiv · GitHub
We only use your address for the digest, and every email has an unsubscribe link. Privacy policy
