Mubit

Attribute decisions and calibrate agents at scale

Mubit links each outcome to the agent decision that caused it, then uses those findings to change how agents handle similar tasks. Attribute decisions across your existing models and calibrate open-weight models with versioned, reversible adapters. Mubit links each outcome to the agent decision that caused it, then uses those findings to change how agents handle similar tasks. Attribute decisions across your existing models and calibrate open-weight models with versioned, reversible adapters. Mubit links each outcome to the agent decision that caused it, then uses those findings to change how agents handle similar tasks. Attribute decisions across your existing models and calibrate open-weight models with versioned, reversible adapters. Mubit links each outcome to the agent decision that caused it, then uses those findings to change how agents handle similar tasks. Attribute decisions across your existing models and calibrate open-weight models with versioned, reversible adapters. Mubit links each outcome to the agent decision that caused it, then uses those findings to change how agents handle similar tasks. Attribute decisions across your existing models and calibrate open-weight models with versioned, reversible adapters.

Get startedRead the docs

01 / How it works

Plug in once. Get back better decisions.

Payment checkout slows down when the service runs out of database connections.

ReflectReuseOutcome feedback Earlier incident payment checkout High latency All DB connections in use Next matching incident payment checkout Latency recovered Resolved with recalled fix Verified fix remembered Mubit Raise connection limit Restart checkout Reflect Reuse Outcome feedback Earlier incident payment checkout High latency All DB connections in use Next matching incident payment checkout Latency recovered Resolved with recalled fix Verified fix remembered Mubit Raise connection limit Restart checkout
See how outcomes connect to decisions
Attributed finding Deploy history was skipped Agent decision Database CPU: 95% Restart database Deploy check skipped Observed outcome Errors return Bad deploy still live Attributed finding Deploy history was skipped Agent decision CPU: 95% Restart database Deploys unchecked Observed outcome Errors return Bad deploy live

02 / Attribute

Find the decision.
Explain the outcome.

Mubit connects each outcome to the decision that caused it, using the run's context and trace. Compare runs to see which choices led to success or failure. Works with any model, including GPT and Claude.

Explore decision attribution
Attributed findingDeploy history was skipped
BaseBefore calibration

Restarts the database when checkout errors spike.

v1Model + adapter

Checks recent deploys first, then rolls back the bad deploy.

Open-weight modelVersioned + reversible

03 / Calibrate

Change the decision.
Improve the next run.

Turn attributed findings into lessons, then into model adapters that change how your agent handles similar tasks. Calibrate open-weight models with versioned updates you can reverse.

Explore calibration

04 / Mubit

Capture. Attribute. Calibrate.

Capture each decision with its context and outcome. Find the decision behind the result, then use that finding to change how your agent handles the next similar task.

Connect the SDK

  1. Capture

    Every decision, its context and its outcome. Three lines of code.

    Capture the run
    Agent run
    Run recordEntry
    ContextDatabase CPU: 95%01
    DecisionRestart database02
    OutcomeErrors return03

  2. Attribute

    Link each outcome to the decision that caused it, across runs.

    Layer 1. Any model, incl. GPT and Claude
    Decision
    Run trace
    Outcome

  3. Calibrate

    Turn the finding into a lesson, then into the model. Versioned and reversible.

    Layer 2. Open-weight models, ~2 min to train
    Lessons
    Adapter

05 / Built for your agent stack

Connect your agent.
Start with one run.

Connect Mubit to your existing agent and start with one workflow. Link the run to a verified outcome so Mubit can attribute the result and build the evidence for calibration.

import os
from dotenv import load_dotenv
import mubit

load_dotenv()

client = mubit.Client(endpoint=os.environ["MUBIT_ENDPOINT"])
client.set_api_key(os.environ["MUBIT_API_KEY"])
client.set_run_id("export-eval-042")

answer = client.recall(
    session_id="exports:checkout-042",
    agent_id="export-eval-agent",
    query="How should checkout recover when the DB pool is full?",
    entry_types=["lesson", "rule"],
)
entry_ids = [entry["id"] for entry in answer.get("evidence", [])]
pip install mubit-sdk python-dotenv
Open the quickstart
SDKsFrameworksCoding agents

Calibrate your agents.
Start with one workflow.

Talk to an engineerRead the docs

Before you build

A few useful details.

What is decision attribution?

Attribution connects an outcome to the agent decision that caused it. Mubit uses the decision's context and run trace to identify what led to that result, across runs. Explore outcome attribution.

What does calibration change?

Calibration turns attributed findings into lessons, then into adapters that change how an open-weight model makes similar decisions. Updates are versioned and reversible. Attribution works across models, including GPT and Claude; adapter calibration is for open-weight models. Read the documentation.

How does Mubit know what worked?

Your application supplies an outcome from a test, an application check or human confirmation. Mubit connects that outcome to the decisions in the run. A completed tool call or a successful HTTP response is not, by itself, evidence that the task succeeded. Report an outcome.

How is this different from storing traces?

A trace records what happened. Mubit connects the result to a specific decision, then uses that finding to change how the agent handles similar tasks on future runs. Explore Mubit.