01
Observe
Traces and metrics from every production run of your models and agents.
runs on: Observe
Auto Research · on Tensile
Recursive self-improvement
is an infra problem.
Auto Research is the research loop that runs on Tensile. It watches how your models and agents behave in production, turns failures into data and environments, trains, evaluates, and ships only what passes. Every turn of the loop is training runs, eval sweeps and rollouts, so the loop is only as fast and as trustworthy as the infrastructure under it.
01The loop
Seven steps, then it starts again.
02
Diagnose
Failures are clustered and scored by verifiers, each in its own sandbox.
runs on: Sandbox
03
Build
New data, rewards and RL environments, made from those failures.
runs on: Sandbox · Registries
04
Train
SFT, DPO or RL runs on your own clusters.
runs on: Training
05
Evaluate
Evals and benchmarks compare every candidate with the model in production.
runs on: Evals
06
Promote
Only candidates that pass are promoted, with the eval evidence kept next to them.
runs on: Registries
07
Serve
The promoted model or agent ships to your endpoints, and the loop starts again.
runs on: Deploy
↻ back to 01 Observe
02Why infra
The loop stalls when the infrastructure does.
Compute-bound
Each turn is GPU work: training runs, eval sweeps and rollouts. Idle or failed GPUs slow the loop down.
Isolated
Agents and verifiers run code you didn't write. Each run gets its own VM.
Evidence first
A change ships only when evals show it is better, and every result traces back to its data, code and model.
Kept running
Niko, our infra agents, debug failed runs and keep your GPUs busy, so the loop keeps turning.
03What researchers run
Start from a question, a paper or a production agent.
Explore a thesis
Scan the literature, then test the idea with A/B variants and a judge model.
Paper → code
Turn a paper into working code, with approval at the steps that matter.
Research sweeps
Run variants across axes and seeds, and compare them on the same benchmark.
Post-train a model
SFT, DPO or RL on a base model and dataset. You get the weights and the metrics.
Improve a live agent
Read a running agent's traces and build the rubrics, verifiers and data to train it.
Code → paper
Turn experiments and results into a write-up: blog, report or paper.
04Start
Auto Research lives inside Tensile.
Sign in to Tensile with your organization's account. Your researchers' runs, evals and models live next to your infra team's clusters, findings and policies.