Skip to main content

​
Overview

Iterate on agents in production without risking all users. Versioning persists configuration snapshots, A/B Testing splits traffic deterministically, and Shadow Mode runs both versions side-by-side.

​
Version Store

Persist agent configuration snapshots to any StorageDriver: 
import { VersionStore, Agent, openai } from "@radaros/core";
import { SqliteStorage } from "@radaros/core";

const store = new VersionStore(new SqliteStorage("versions.db"));

// Save a version
const version = await store.save({
  agentName: "support-agent",
  instructions: "You are a customer support agent...",
  modelId: "gpt-4o",
  providerId: "openai",
  toolNames: ["search_kb", "create_ticket"],
  temperature: 0.7,
});

// List all versions
const versions = await store.list("support-agent");

// Compare versions
const diff = store.diff(versions[0], versions[1]);
// [{ field: "modelId", before: "gpt-4o", after: "gpt-4o-mini" }]

​
A/B Testing

Split traffic between control and variant agents: 
import { ABRouter } from "@radaros/core";

const router = new ABRouter({
  name: "new-instructions-test",
  control: { agentName: "support-v1" },
  variant: { agentName: "support-v2" },
  trafficSplit: 0.1,         // 10% to variant
  routing: "user",           // deterministic by userId
  autoRollback: {
    errorRateThreshold: 0.15, // rollback if >15% errors
    windowMs: 300_000,        // in last 5 minutes
  },
});

// Route a request
const variant = router.route({ userId: "user-123" });
// Returns "control" or "variant" deterministically

// Record outcome
router.recordRun(variant, true, 1200, 500);

// Check metrics
const metrics = router.getMetrics();
// { control: { totalRuns: 900, successCount: 880, ... },
//   variant: { totalRuns: 100, successCount: 95, ... } }

// Auto-rollback check
if (router.shouldAutoRollback()) {
  console.log("Variant error rate too high, rolling back");
}

​
Routing Strategies

StrategyBehavior
"random"Pure random split
"user"Hash userId for deterministic assignment
"session"Hash sessionId for per-session consistency

​
Shadow Mode

Run both agents in parallel, return primary result, log comparison: 
import { ShadowRunner, Agent, openai } from "@radaros/core";

const shadow = new ShadowRunner(
  primaryAgent,
  shadowAgent,
  {
    compareOutputs: (primary, shadow) => ({
      match: primary.text === shadow.text,
      similarity: computeSimilarity(primary.text, shadow.text),
      differences: ["Tool call count differs"],
    }),
  },
);

const result = await shadow.run("How do I reset my password?");
// result.primaryOutput — always returned to user
// result.shadowOutput  — logged for comparison
// result.comparison    — { match, similarity, differences }

​
Events

EventPayload
version.created{ agentName, versionId }
ab.routed{ testName, variant, userId }
ab.metrics{ testName, control, variant }
shadow.compared{ agentName, match, similarity }