Context
State, user, environment, and operational constraints.
env.observe()Reinforce OS
The operating system for adaptive decision systems.
Build applications that experiment, infer causality, and improve policies over time with experimental design, causal inference, and reinforcement learning.
Reinforce OS loop
A closed loop for learning systems.
Reinforce OS turns experimentation into infrastructure. Every decision becomes part of a feedback loop: observe context, choose an action, measure the outcome, estimate causal effects, and refine the policy.
Decision
Choose an action using the current policy and safety guardrails.
policy.select()Intervention
Assign treatment, workflow, prompt, price, or recommendation.
decision.actionOutcome
Measure what happened and preserve the audit trail.
observe(outcome)Inference
Estimate causal effects, uncertainty, and counterfactuals.
posterior.update()Policy Update
Refine future decisions without losing track of exploration.
policy.deploy()Core primitives
The primitives of adaptive software.
Reinforce OS exposes the basic building blocks required to turn applications into learning systems.
Environments
Define the system, context, constraints, and observable state.
reinforce.environment("onboarding")Actions
Represent interventions the system can choose.
actions: ["guided_setup", "video_intro"]Policies
Decision rules that map context to action.
policy.select({ state, constraints })Rewards
Objectives, tradeoffs, and delayed outcomes.
reward: "activated_within_7_days"Experiments
Randomization, assignment, power, crossover designs, and adaptive sampling.
strategy: "bandit"Counterfactuals
Estimate what would have happened under alternative actions.
estimate.do(action_alt)Architecture
Built as a learning-system stack.
Layered infrastructure for experiments, causal inference, adaptive policies, and production data.
Application Layer
Personal apps, enterprise workflows, education, agent systems.
Reinforce OS API
Environments, actions, policies, rewards, observations.
Experiment Core
RCTs, crossover designs, randomization, power, compliance.
Causal Engine
DAGs, SCMs, counterfactuals, Bayesian posteriors.
Adaptive Layer
Bandits, policy gradients, online RL, safe exploration.
Data Layer
Events, outcomes, metrics, audit trails, model inputs.
Code / API
Developer-first by design.
Define a decision problem, deploy a policy, observe outcomes, and let the system learn.
import { reinforce } from "@dooperator/reinforce";
reinforce.configure({ apiKey: process.env.REINFORCE_API_KEY });
const env = reinforce.environment("onboarding");
const policy = env.policy({
strategy: "bandit",
actions: ["guided_setup", "video_intro", "blank_canvas"],
reward: "activated_within_7_days",
});
const decision = await policy.select({
userId: user.id,
state: {
role: user.role,
experience: user.experience,
teamSize: user.teamSize,
},
});
await app.show(decision.action);
await policy.observe({
decisionId: decision.decisionId,
outcome: {
activated: true,
satisfaction: 9,
},
});sample output
{
"action": "guided_setup",
"confidence": 0.74,
"exploration": true,
"policyVersion": "π_0042"
}Methodology
Three engines. One adaptive loop.
Reinforce OS integrates experimental design, causal inference, and reinforcement learning into one production workflow.
01
Controlled exploration.
RCTs, crossover trials, stratified randomization, power calculations, and pre-registered hypotheses are built into the workflow.
- Parallel and crossover RCTs
- Adaptive randomization
- Sample size and power
- Pre-registration
- Compliance checks
02
Effects, not correlations.
Structural causal models, DAGs, counterfactuals, and Bayesian posteriors turn observations into defensible causal estimates.
- DAG modeling
- ATE and HTE estimation
- Sensitivity analysis
- Credible intervals
- Counterfactual simulation
03
Policies that improve.
Bandits and online reinforcement learning can update policies as evidence accumulates while respecting exploration and safety constraints.
- Thompson sampling
- UCB
- Contextual bandits
- Online policy learning
- Safe exploration
Agent policy testing
Run A/B tests between agent prompt versions.
Define two policy versions as experiment conditions, log outcome metrics after each agent run, and let Bayesian analysis tell you which policy wins.
01
Define two policies.
Create an experiment with baseline_policy and new_policy as conditions. Any domain: support, sales, ops, content.
02
Log business outcomes.
After each agent run, POST the condition used and the observed metrics — resolution rate, escalation rate, override rate, cost.
03
See which policy wins.
Bayesian analysis accumulates as data arrives. P(new_policy better) updates in real time. Share results with a link.
Applications
One platform. Many adaptive applications.
DoOperator builds the platform and the applications on top of it so the infrastructure stays grounded in real decision loops, including enterprise agent workflows.
SteadyPractice
Personal science for sleep, focus, habits, recovery, and behavior change.
Open →Decision Process
Experimentation and adaptive workflows for teams, operators, and product systems.
Open →DoOperator Education
The open curriculum covering experimental design, causal graphs, Bayesian inference, bandits, and reinforcement learning.
Open →DoOperator Research
21,000+ curated papers on causal inference, experimental design, heterogeneous effects, bandits, and reinforcement learning — with deep wiki summaries.
Open →Agent policy A/B
Run controlled experiments between AI agent prompt versions, models, or routing policies. Measure resolution rate, escalation rate, and override rate.
Open →DoOperator Education
Courses and education built into the platform.
DoOperator Education teaches the theory behind Reinforce OS: experimental design, causal graphs, Bayesian inference, bandits, reinforcement learning, and policy evaluation.
Technical foundation
Rigorous by construction.
Most decision software is a correlation engine with a dashboard. Reinforce OS treats interventions, counterfactuals, confounders, uncertainty, and policy updates as first-class system objects.
RCT
Randomized assignment and controlled experimentation.
SCM
Structural causal models for interventions and counterfactuals.
MAB
Bandit allocation for adaptive exploration.
OPE
Off-policy evaluation before deploying new strategies.
RL
Online policy improvement under uncertainty.
Mission
Replace guesswork with learning systems.
Decisions are everywhere: what to recommend, what to teach, what workflow to show, what intervention to try, what policy to deploy. Today most of those decisions are static, intuitive, or optimized against weak correlations.
Reinforce OS makes them adaptive. It gives builders the infrastructure to test interventions, estimate causal effects, manage uncertainty, and improve policies over time.
Final CTA
Build the next generation of adaptive applications.
We are looking for technical collaborators, early customers, researchers, and investors who believe software should learn from experience.