Statsig
Type: low-code · Vendor: Statsig (OpenAI) · Language: TypeScript · License: proprietary · Status: active · Status in practice: mature · First released: 2020
Statsig is a feature-management and experimentation platform that runs feature gates, controlled rollouts, and bandit-driven experiments over production traffic.
Description. Statsig provides feature gates that expose features to a configurable percentage of users and can be turned off in production through an emergency disable switch. Its Autotune product runs multi-armed bandit experiments that use Thompson Sampling to allocate traffic toward better-performing variants. The platform is general-purpose experimentation and feature management rather than agent-specific.
Agent loop shape. Statsig is not an agent loop; it is a decisioning layer queried at runtime. Application SDKs check a feature gate or experiment for a given user and Statsig returns whether the gate is on and which variant the user is assigned. For Autotune experiments, the bandit reallocates traffic across variants over time based on observed rewards, and operators can flip a gate off out-of-band to disable a feature in production.
Primary use cases
- feature gates with percentage-based rollouts
- emergency disabling of features in production
- bandit-driven variant experimentation
Key concepts
- Feature gate → kill-switch (docs) — A runtime-checked switch that decides whether a code branch is on for a given user; used for percentage rollouts and as an emergency disable switch.
- Experiment (docs) — A controlled A/B test that assigns users to variants and measures primary and secondary metrics against control with confidence intervals to test a hypothesis.
- Autotune → bayesian-bandit-experimentation (docs) — A multi-armed bandit experiment that uses Thompson Sampling to continuously reallocate traffic toward the best-performing variant instead of holding a fixed split.
- Layer (docs) — An allocation construct that partitions traffic so concurrent experiments do not interfere; an experiment runs in its own layer by default.
Patterns this low-code implements —
- ★Bayesian Bandit Experimentation
Statsig Autotune is a multi-armed bandit that uses Thompson sampling to dynamically reallocate traffic toward better-performing variants in real time, minimizing regret rather than running a fixed-sp…
- ★Kill Switch
A Statsig feature gate doubles as an out-of-band emergency disable switch that immediately turns off a code branch for users in production without a redeploy; note this is general feature management…
- ★★Shadow Canary
A Statsig feature gate can be set to expose a new code branch to only a small percentage of the user base, acting as a canary rollout that limits blast radius before a feature reaches everyone; this…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
- full-code · framework
Braintrust
Braintrust is an LLM eval/experiment platform; Statsig brings the general-purpose feature-flag and online-experimentation layer that an LLM team would pair with eval tooling.
- low-code · visual builder
Langfuse
Langfuse handles LLM tracing and evaluation while Statsig handles production feature gating and bandit experiments; they cover adjacent halves of the rollout/measurement loop.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: partial