The Intelligence Improvement StackBuild AI That Gets Better

Test AI models on real tasks, measure how they perform, and improve them through continuous evaluation and optimization.

START IMPROVING
EXPLORE THE LAB

$ pip install gradientpip install gradient

Gradient Raises $400K Pre-Seed

Backed by angels and operators from OpenAI, Anthropic, DeepMind, NVIDIA, Scale AI, xAI, and Meta AI.

Evaluate and improve leading models through one platform.

Lab. Post-train self-improving agents

Start Training

FIG.1

Evaluation workflow: task, run, score, improve around an AI model
01Evaluation Environments

Turn real-world tasks into repeatable tests for AI agents. Measure performance, find failures, and improve what matters.

01.1Built for coding, research, browser, and tool-use tasks
01.2Test any model against the same tasks and scoring criteria
01.3Turn every failure into a measurable path to improvement

FIG.2

Model evaluation benchmark: select, run, score, and track model performance
02Evaluations

Benchmark AI models on real tasks and see exactly where they succeed or fail.

02.1Compare leading open and closed models
02.2Run evaluations with no infrastructure to manage
02.3Track scores, failures, and improvements over time

FIG.3

Agent optimization loop: create variants, test, measure, and keep the best performer
03Optimization

Continuously improve AI agents against the tasks that matter.

03.1Create and test multiple agent variants
03.2Measure every change against the same benchmark
03.3Automatically keep the best-performing version

FIG.4

Deployment pipeline: evaluate, deploy, and monitor agents in production
04Deployment

Move your best-performing agents from evaluation into production.

4.1Deploy the highest-scoring agent version
4.2Monitor performance on real-world tasks
4.3Turn production failures into the next improvement cycle

“Like a human, our agents learn for themselves to achieve successful strategies that lead to the greatest long-term rewards.”

Google

David Silver

Google DeepMind

Gradient Environments

A library of real-world tasks for evaluating and improving AI models.

EXPLORE ENVIRONMENTS
Explore
My Stars
My Environments
Featured9
Show All
gradient
2

code-repair

Fix broken code against a test suite and scoring rubric...

codingrepair+1
Updated 8 days ago
v0.3.8
gradient
6

browser-nav

Navigate web apps and complete multi-step browser tasks

browsernav+1
Updated 11 days ago
v0.2.5
stochi0
3

research-qa

Answer research questions with cited sources and scoring...

researchqa+4
Updated 2 months ago
v0.2.0
Gradient3
gradient
8

mini-swe-agent-plus

Mini SWE Agent Plus environment for solving SWE issues inside Pri...

swesandbox+1
Updated 3 days ago
v0.2.23
gradient
6

deepdive

DeepDive QA RL environment with a Serper-powered search tool

rlqa+1
Updated 11 days ago
v0.2.5
gradient
3

science-env

A collection of challenging single-turn science problems

sciencesingle-turn
Updated 11 days ago
v0.1.3
Evals13
Show All
hud
18

hud-text-2048

Text-based 2048 game for training agents to reach target tiles through strategic moves

gametext+2
Updated 7 months ago
v0.1.0
hud
18

hud-text-2048

Text-based 2048 game for training agents to reach target tiles through strategic moves

gametext+2
Updated 7 months ago
v0.1.0
will
29

tool-agent

Tool-use agent environment with API, SQL, and search tools

toolsagent+2
Updated 2 months ago
v0.1.0
Environment SDK
123456789
from gradient import Environment
env = Environment(
        name="code-repair",
        tasks=task_list,
        tools=[search, edit, run],
        scorer=pass_at_k,
    )

Build tasks, tools, and scoring rules for any workflow.

Evaluation Engine
$ gradient eval qwen3 \
  --env code-repair \
  --compare gpt-4o,gemini \
  --track scores

Run models against the same tasks and measure performance.

Optimizer
A

variant-a

score: 71%

B

variant-b

score: 84%

C

variant-c

score: 78%

Test agent variants automatically and keep the best-performing version.

Optimization. Turn evaluation results into
better-performing AI agents.

START IMPROVING

Run agents against real tasks, identify where they fail, test improvements, and continuously promote the versions that perform best.

Automatic Optimization
Generate and test new agent configurations across prompts, tools, reasoning strategies, and model settings.
Version Comparison
Compare every new version against the same evaluations to see exactly what improved and what regressed.
Continuous Improvement
Keep your best-performing agent and continuously improve it as new tasks, failures, and evaluation data appear.

Turn every failure into the next improvement.

Capture failed tasks, turn them into new evaluations, and use the results to build agents that become more accurate, reliable, and capable over time.

Open Source. Build on Gradient

DISCOVER

Open Source

Explore the code behind Gradient

Open-source training and evaluation framework, adapted for testing, benchmarking, and improving AI agents.

VIEW ON GITHUB

Latest Articles.