Self-improving AI model
A model that gets better at your work.
Sensei learns from verified agent outcomes to deliver higher accuracy with fewer human corrections and lower token costs. Its routing shares what your agents learn across the models you use.
SenseLab benchmark
Sensei vs.
Opus 5.5 alone
93%fewer human handoffs
70%+lower estimated token costs
SenseLab benchmark · Support and CI-repair workflows
View the benchmark
Support-ticket and CI-repair streams across ten seeds, including changes to workflow rules during evaluation.
Human handoffs: 82 for Sensei vs. 1,215 for Opus 5.5 alone per 6,000 tasks, a 93.25% reduction rounded to 93%.
Estimated token cost per 10,000 tasks: $14 for Sensei vs. $151 for Opus 5.5 alone, a 90.73% reduction. Estimates use projected serverless costs and proposed Sensei pricing.
These results apply to this evaluation. They compare model performance, rather than savings from model routing.
Watch how Sensei works.
1:15 · Product overview
See how agent outcomes become shared learnings and custom models.
Video not loading? Open the video directly.
Read the video summary
Agents can repeat mistakes when business rules and workflows change. Manually fine-tuning a model or putting the full task history into each prompt takes ongoing work.
Sensei captures agent outcomes, learns from what worked and what failed, and uses those learnings to guide agents during execution. Fine-tuning turns verified work into specialized models for different use cases.
The video illustrates learning across support, coding, operations and billing agents. It shows project-specific model versions improving as work continues, and connecting Sensei to existing agent tools.
Its benchmark compares Opus 5.5 before and after adding Sensei on tasks after workflow rules changed: 0% versus 96% correct. Other figures compare Sensei’s own model with Opus 5.5. These are SenseLab evaluation results, rather than a guarantee for every workload.
01 / Self-improving model
Build expertise into the model you run.
Sensei turns task outcomes into reusable learnings, fixes failed approaches through recursive self-improvement, and fine-tunes models for your projects and use cases.
Task resultsScored learnings
Run
Sensei captures your task results, scores and weights each outcome, and automatically labels what your agents learned.
02 / Model routing
Route for accuracy and cost. Share what works.
Sensei’s router selects models for each workload’s accuracy and token cost requirements. Its learning engine turns verified outcomes into shared learnings that other agents can use, including those running different models.
What works in one run becomes available to the next agent.
Explore a model choice
IllustrativeShared learnings
Carried across- Learning
- Check eligibility before issuing a refund.
- Evidence
- Verified task outcomes.
- Next
- Reuse the procedure on the next ticket.
Run the next ticket on Sensei. Verified results update the shared learnings available to other agents.
Custom models
A custom model for every project.
Fine-tune Sensei on your project data, workflows and verified agent outcomes. Build separate models for different projects, with training and evaluation managed by SenseLab.Available with Teams.
Your existing workflow
Run Sensei where your agents already work.
Use Sensei as the model in Codex, Cursor or Claude Code. Keep your existing tools and workflows while Sensei learns from task outcomes.
- Codex
- Cursor
- Claude Code
Before you join
A few practical questions.
What does “self-improving” mean?
Sensei uses task outcomes to guide execution and build training data for future model versions. Learning from outcomes and training model weights operate on different cycles.
Can I use other models alongside Sensei?
Yes. Route workloads based on accuracy and token cost. Agents using those models can contribute to and benefit from the shared learnings inside Sensei’s router.
Can I train models for different projects?
Yes. Teams can fine-tune separate models on project-specific data, workflows and verified agent work. SenseLab manages training and evaluation.
What happens after I join?
We’ll save your details and interests, then email you when Sensei launches with information about access and next steps.
Will my project’s work train a shared model?
The proposed policy is Dev sharing on by default with a Privacy Mode option; Teams sharing only if an admin opts in; no Enterprise data in base-model training. We’ll confirm the controls and terms before you connect a project.
The Sensei preview
Self-improving models.
Fine-tuned for your team.
Train Sensei on your data, workflows and verified agent outcomes. Create a model for each project, with training and evaluation managed by SenseLab.