Split is a feature flag and experimentation platform founded in 2015 in Redwood City by Adil Aijaz, Patricio Echagüe, and Trevor Stuart. In May 2024 Harness announced its acquisition of Split (deal closed June 11, 2024), and the product was rebranded as Harness Feature Management & Experimentation (FME). Split is now one product inside Harness's broader CI/CD and AI-delivery platform alongside Continuous Delivery, Continuous Integration, Cloud Cost Management, and AI-powered code agents.
The sections below cover how the platform works, what it is good at, and where it sits relative to Confidence, the experimentation platform Spotify has run for 15 years.
How does Split (Harness FME) work?
Harness FME runs as a managed service that sits between your application code and your decision data. Application code calls Harness FME SDKs to evaluate feature flags, with the SDK maintaining a cached configuration that updates over a streaming or polling connection so flag evaluation is local. The same SDKs log experiment exposures: when a user is assigned to a treatment, the SDK records that assignment to be joined later with metric data.
For experimentation, the stats engine ships frequentist hypothesis testing, mSPRT (mixture sequential probability ratio test) for sequential testing, fixed-horizon analysis, sample ratio mismatch detection (chi-squared with a p<0.001 threshold), guardrail metrics, and Multiple Comparison Correction (MCC). CUPED variance reduction, which uses pre-experiment data to tighten confidence intervals, is not listed in the public stats documentation as of 2026.
Since the acquisition, Harness FME has added Warehouse-Native Experimentation (analysis runs against the customer's data warehouse), AI-driven experiment summarization, an MCP server for AI IDEs, rule-based segments, and a Reallocate Traffic API. The March 2026 Harness platform release added AI-delivery release coordination across the broader platform.
What Split (Harness FME) is good at
Harness FME's strengths sit in two places: the experimentation methodology Split shipped pre-acquisition, and the broader Harness platform integration that came with the acquisition.
- mSPRT sequential testing. A peeking-safe statistical method that lets you stop experiments early without inflating false- positive rates. mSPRT is a specific family of sequential tests that some practitioners prefer over Group Sequential Tests for the always-valid guarantee shape it produces.
- Sample ratio mismatch detection with a chi-squared p<0.001 threshold. SRM detection flags traffic split anomalies that usually indicate a bucketing bug.
- Guardrail metrics and Multiple Comparison Correction. Standard rigor-by-default surfaces shipped on paid tiers.
- Warehouse-Native Experimentation. Analysis runs against the customer's data warehouse rather than Harness's storage, added post-acquisition.
- Harness platform integration. CI/CD pipelines, deployment automation, cloud cost management, AI code agents, and (since March 2026) AI-delivery release coordination under one vendor. For teams that have standardized on Harness for the broader software-delivery stack, Harness FME slots into that platform.
- MCP server for AI IDEs. Tied to the broader Harness AI-coding integrations.
- Free Developer tier. Up to 10 seats; usage-based Growth tier; sales-gated Enterprise.
- Customer base. Twilio, Salesforce, GoDaddy, Electronic Arts, Rocket Mortgage, WePay, Healthfirst, all carried through from Split to Harness FME marketing.
For organizations that want experimentation alongside CI/CD, release coordination, and AI-driven code workflows under one vendor, Harness FME is the integrated answer.
Confidence is the platform Spotify uses to decide what its product becomes. The defaults reflect 15 years of running experiments at scale, including the failure modes that only show up at scale. It is now available to teams outside Spotify.
Where Confidence and Split diverge
The largest difference is vendor parent. Confidence is built and operated by the team that runs Spotify's experimentation platform; the roadmap is set by that team. Harness FME's roadmap is set inside Harness, whose primary business is CI/CD, release coordination, and AI-driven software delivery. Experimentation is one product line inside a platform whose other products (Continuous Delivery, Continuous Integration, Cloud Cost Management, AI code agents) compete for engineering investment.
The same Confidence platform serves 300+ Spotify teams running 10,000+ experiments per year across 750M users in 186 markets. 42% of those experiments are rolled back after guardrail metrics flag a regression. Split has 11 years of commercial history with public references including Twilio, Salesforce, GoDaddy, Electronic Arts, and Rocket Mortgage; Harness FME inherits that customer base and is now sold inside the broader Harness platform's go-to-motion.
Methodology specifics differ. Confidence's CUPED uses the Negi–Wooldridge full regression estimator. Harness FME does not list CUPED in its public stats documentation as of 2026. Both products ship sequential testing (Confidence: Group Sequential Tests with always-valid inference; Harness FME: mSPRT), sample ratio mismatch detection, and guardrail metrics. For buyers who specifically want CUPED at Spotify scale on a managed warehouse- native platform, Confidence is the focused option; for buyers who want mSPRT-based sequential testing inside a CI/CD platform, Harness FME is the natural fit.
OpenFeature integration: Confidence's iOS and Android OpenFeature provider SDKs were donated to the CNCF (Cloud Native Computing Foundation), and Spotify holds a seat on the OpenFeature governance committee. Split (Harness FME) maintains official OpenFeature providers across .NET, Java, JavaScript, and Go but is not on OpenFeature governance.
Each platform fits a different organization
Harness FME fits organizations that have standardized on Harness for CI/CD and want experimentation alongside their build, deploy, and release coordination workflows under one vendor. The platform-bundle economics and the AI-delivery release coordination story are the selling points.
Confidence fits teams that want experimentation as a single managed product, with opinionated defaults built on 15 years of Spotify operating evidence and a roadmap set by the team that built it. The methodology depth and the operating-history evidence are the selling points.
The choice is about which vendor parent sits behind your experimentation program for the next five years. A free self-serve trial of Confidence is available at confidence.spotify.com without going through procurement. The Confidence vs Split head-to-head covers vendor parent, methodology, and platform integration in detail. For teams already on Split who want to know what other options exist post-acquisition, see Top 7 alternatives to Split.