The bottleneck moved
Building got cheap. A change that used to take a sprint now takes an afternoon, and agentic tools keep pushing that cost down. What didn’t get cheap is knowing what each change did — what to keep, what to revert, what to double down on. If you can ship ten changes a week but evaluate two, the other eight are guesses. Traffical is built on the premise that three things have to keep pace with each other:- how fast you ship user-facing changes,
- how fast you detect and contain the ones that hurt,
- and how fast you produce evidence you trust.
Parameters, not feature flags
Most tools in this space start from the feature flag — a boolean, on or off. Useful, but a ceiling: real products tune prices, adjust algorithm weights, pick headlines, and configure ranking rules, not just toggle features. Traffical starts from the parameter: a typed, versioned value with a default. Experiments, rollouts, feature flags, and adaptive optimization are all policies that control how parameter values are assigned. Because the primitive is a value rather than a switch, an A/B test on a headline, a bandit on a price point, and a gradual rollout of an algorithm change all run through the same system — same SDKs, same layers, same dashboard.- Feature flag approach
- Traffical approach
Local resolution
Traditional experimentation platforms decide on a server and serve the answer over the network — a round trip per request. Traffical resolves locally: your SDK fetches a pre-built config bundle and every decision happens in-process, in sub-milliseconds. No per-request API calls. Works offline, works at the edge, and if Traffical is unreachable your app keeps running on the last known configuration.Concurrent experiments that don’t collide
Running one experiment is straightforward. Running fifty at once, across the same users, without them contaminating each other is the real problem. Traffical organizes parameters into layers: within a layer, a user is in at most one policy; across layers, bucketing is orthogonal, so experiments don’t interfere. Teams run in parallel without coordinating calendars.Measurement where your data already lives
You don’t need a second event pipeline to measure outcomes. Define metrics as SQL against your own warehouse — Postgres, BigQuery, Snowflake, Databricks, ClickHouse — and Traffical executes the queries, joins assignments to outcomes, and produces the same per-allocation statistics as its native events. Outcomes SDKs never see — renewals, refunds, support tickets — count the same as clicks. Use warehouse-native metrics, Traffical-native events, or both at once.Governance built in — for humans and agents
Speed without control is how incidents happen, and more of your changes are shipped by agents every quarter. Traffical’s governance surface, changes, makes the safe path the default: every change states its intent before it carries traffic, every traffic-bearing phase requires an approved measurement plan, and every transition writes a decision record — who acted, what the evidence said, what changed. A computed risk class scales the friction. A low-risk change with certified measurement coverage starts without ceremony; a pricing change on a backend surface waits for human sign-off. Agents act through the MCP server under the same gates, not around them: safety actions like pause and revert are always cleared, while promoting a winner into the product default is always a human checkpoint. See approvals and autonomy for the full model.From static tests to adaptive optimization
A/B testing answers “which variant is better, on average?” — valuable, but manual and one-size-fits-all. When the best variant shifts over time, or differs per user, adaptive policies move traffic toward better performers automatically: Thompson Sampling, UCB1, Epsilon-Greedy, and linear contextual bandits that personalize per user. See optimization and choosing an algorithm.Developer-first
- Type-safe SDKs for TypeScript (Node, browser, React, Svelte, React Native), Python, PHP, and Swift — one shared conformance spec, so a given unit buckets identically on every platform
- Config as code: parameters, events, and metrics in version-controlled YAML, with
push,pull, and astatuscheck that gates CI on drift - A visual editor for no-code UI experiments, and DevTools to inspect SDK state and override parameters live
- An MCP server and agent skill so coding agents can run experiments under the same governance you do