autograd

Inference for agent systems

Inference that improves from every run.

Start with Qwen 3.8 27B. Connect your agent once, capture real trajectories and reward signals, then evaluate, reinforce, and promote better checkpoints.

Read the quickstart See the learning loop
The Autograd reinforcement learning loop An agent request moves through inference to an outcome and reward, then a gradient update produces the next model version. AGENT INFERENCE OUTCOME REWARD ∂L / ∂θ forward pass backward pass
01
01Start with inference

A useful model
on the first call.

Inference is not a side feature. It is the beginning of the data loop. Every request establishes the context needed to understand how an agent behaves in production.

Initial model Qwen 3.8 27B

Built for agent inference and exposed through a familiar client interface. Change the endpoint, keep the agent, and start building a history of the work it actually performs.

Interface
OpenAI-compatible
Workload
Agent inference
Access
Private preview
Interface preview

Your agent already knows how to connect.

The first API surface follows familiar OpenAI client semantics. Endpoint details and the public model slug will be finalized with the inference release.

Open the quickstart
python 
from openai import OpenAI

client = OpenAI(
    base_url="https://api.autograd.run/v1",
    api_key="AUTOGRAD_API_KEY",
)

response = client.chat.completions.create(
    model="qwen-3.8-27b",
    messages=[{
        "role": "user",
        "content": "Plan the task."
    }],
)
02The reinforcement loop

The run is
the training data.

Autograd joins inference with the outcomes that happen after inference. The result is a controlled path from production behavior to reinforcement learning.

  1. 1

    Run

    Serve the agent through Autograd inference. Capture model steps, tools, latency, usage, and model version.

  2. 2

    Reward

    Attach product outcomes, user corrections, graders, or explicit scores through the SDK.

  3. 3

    Reinforce

    Evaluate and train against the work your agent actually performs—not a generic benchmark alone.

  4. 4

    Promote

    Compare the candidate with the baseline and deliberately promote a better checkpoint.

03Add the missing signal

The Autograd SDK

Inference sees the call.
Your product knows the outcome.

A response cannot tell us whether the user accepted it, whether the tool succeeded, or whether the task was completed. The SDK closes that gap.

  • User signalaccept · retry · edit
  • Agent signaltool success · task complete
  • Learning signalreward · grader · policy
sdk preview 
with autograd.trace() as run:
    result = agent.invoke(task)

    run.reward(
        "task_success",
        1.0,
    )

    run.event(
        "user.accepted",
        {"surface": "copilot"},
    )

Proposed SDK surface. Names may change before release.

04Optimize deliberately

Nothing learns blindly.
Nothing ships accidentally.

01

Choose what trains

Select trajectories and define which data is permitted to influence a model update.

02

Evaluate the change

Compare every candidate against a baseline and a held-out evaluation set.

03

Promote with intent

Keep checkpoint history visible and require an explicit decision before production traffic moves.

05Build status

Private preview

Built in the open,
described precisely.

First release

Qwen 3.8 27B inference, OpenAI-compatible client surface, structured run capture.

In development

SDK reward events, trajectory browser, dataset selection, baseline evaluations.

Coming next

Automated RL runs, candidate evaluation, checkpoint comparison, promotion and rollback.

One endpoint. A compounding system.

Put your agent
on a learning path.

Explore the docs