Config Targeting

Configure targeting rules for configs to control which variations serve to different contexts. Works the same for both completion and agent mode.

Prerequisites

• LaunchDarkly account with AgentControl enabled
• The LaunchDarkly MCP server, connected as a member who can change config targeting in the environment
• Project key and environment key
• Existing config with variations (use launchdarkly-configs-create skill)

API Access

Use the LaunchDarkly MCP server tools: they authenticate with the connection the customer set up. If a step has no MCP tool, tell the user which change to make in the LaunchDarkly UI instead. Never ask the user for an API access token.

Tools

update-agent-config with resource: "targeting" changes what live traffic receives and requires confirm: true. Before calling it, tell the user what the config serves now and what will change, and pass confirm: true only after they agree.

Core Concepts

Evaluation Order

Targeting rules evaluate in this order (same as feature flags):

1. Individual targets - Specific context keys (highest priority)
2. Custom rules - Attribute or segment conditions, top to bottom; the first match wins
3. Default rule - Fallthrough for all others

When an environment's targeting is off (on: false), every context gets the off variation instead. A new config's default rule serves an auto-generated disabled variation, so to start serving a variation, change the default rule. The turnTargetingOn instruction does not work for configs; use the on field.

Key Concepts

• variationKey: fallthrough and targetingInstructions take a variationKey: a variation key or name from get-agent-configs. You don't need variation UUIDs or indexes.
• Percentages: a fallthrough split takes whole percentages that sum to 100.
• Clause logic: Multiple clauses = AND, multiple values = OR
• Null attributes: Rules with null/missing attributes are skipped
• Approvals: When the environment requires approval, update-agent-config makes no change and returns requiresApproval: true with the instructions to submit (Step 5).

Workflow

Step 1: Read the Current Targeting

Call get-agent-configs with projectKey, configKey, section: "targeting" and env. Note the variation keys and names, each rule's _id and its clauses' _id values, the individual targets, and what the default rule serves. Call get-agent-configs with only projectKey and configKey to see what each variation contains.

Step 2: Edit the Default Rule

Call update-agent-config with resource: "targeting" and a fallthrough:

{
  "projectKey": "my-project",
  "configKey": "chat-assistant",
  "resource": "targeting",
  "env": "production",
  "fallthrough": {"variationKey": "v1-gpt4o"},
  "confirm": true
}

• To serve one variation to every context no rule matches: "fallthrough": {"variationKey": "v1-gpt4o"}.
• To split that traffic: "fallthrough": {"weights": [{"variationKey": "v1-gpt4o", "percent": 80}, {"variationKey": "v2-claude", "percent": 20}]}.

Important: A new config's default rule serves an auto-generated disabled variation. Until you point the default rule at a real variation, the SDK returns enabled=False to every context.

Tell the user what the default rule served before (from Step 1) and what it serves now, so they can undo the change.

Step 3: Add Targeting Rules

Call update-agent-config with resource: "targeting" and one or more targetingInstructions. The rule kinds are addRule, removeRule, addClauses, removeClauses, reorderRules, replaceRules, updateRuleDescription, addValuesToClause, removeValuesFromClause and updateClause. Take ruleId and clauseId from the _id values in Step 1.

Attribute-based rule:

{
  "projectKey": "my-project",
  "configKey": "model-selector",
  "resource": "targeting",
  "env": "production",
  "targetingInstructions": [{
    "kind": "addRule",
    "variationKey": "sonnet",
    "clauses": [{
      "contextKind": "user",
      "attribute": "selectedModel",
      "op": "contains",
      "values": ["sonnet"],
      "negate": false
    }]
  }],
  "confirm": true
}

Segment rule (the instruction only):

{
  "kind": "addRule",
  "variationKey": "experimental",
  "clauses": [{
    "contextKind": "",
    "attribute": "segmentMatch",
    "op": "segmentMatch",
    "values": ["beta-testers"],
    "negate": false
  }]
}

Leave contextKind blank for segment clauses.

To change which variation an existing rule serves, send replaceRules with every rule in order, each with its variationKey and clauses. replaceRules overwrites all rules, so build the list from the rules Step 1 returned.

Each rule serves one variation. For a percentage split inside a rule, or a progressive or guarded rollout of a config, tell the user to make the change in the LaunchDarkly UI.

Step 4: Target Individual Contexts

Use targetingInstructions with addTargets, removeTargets, clearTargets or replaceTargets. addTargets and removeTargets require contextKind and values:

{
  "projectKey": "my-project",
  "configKey": "chat-assistant",
  "resource": "targeting",
  "env": "production",
  "targetingInstructions": [{
    "kind": "addTargets",
    "variationKey": "v1-gpt4o",
    "contextKind": "user",
    "values": ["user-key-1"]
  }],
  "confirm": true
}

Individual targets override every rule.

Step 5: Submit for Approval When Required

If update-agent-config returns requiresApproval: true, the change wasn't made. Ask the user who should review it; if they name a person, find their member ID with search-members. Then call create-agent-config-approval with projectKey, configKey, env, a one-line description of the change (required), at least one reviewer in notifyMemberIds or notifyTeamKeys, and the returned instructions. Show the user the request's approvalRequest.url: a person approves or declines it in LaunchDarkly, and you can't. When the user says it has been reviewed, check it once with get-agent-config-approvals and the request's id. Only when reviewStatus is approved, apply it with apply-agent-config-approval (id and confirm: true). If it is pending or declined, tell the user and don't apply it.

Step 6: Verify

Call get-agent-configs with section: "targeting" again and confirm that the default rule, rules and individual targets match the request. section: "status" shows the config's status in every environment, and section: "health" reports missing variations, models, prompts and tool references.

Roll Back a Config

When a variation misbehaves, for example after a model or prompt change:

1. Call get-agent-configs with section: "targeting" and the environment to find what serves the variation: individual targets, a rule or the default rule.
2. Find the variation that served before with search-change-history (resourceKind: "aiconfig", resourceKey set to the config key, and the environment). If the history doesn't show it, ask the user which variation to restore.
3. If the default rule serves it, call update-agent-config with resource: "targeting" and "fallthrough": {"variationKey": "<earlier variation>"}. If a rule or individual targets serve it, change that rule (Step 3) or those targets (Step 4). If no earlier variation is safe to serve, turning targeting off with on: false serves the off variation to every context in that environment.
4. Verify as in Step 6, and tell the user what you changed.

If the variation itself was edited (its prompt or model changed, not which variation serves), restore its earlier version as described in launchdarkly-configs-update.

Operators Reference

Clause Structure

{
  "contextKind": "user",
  "attribute": "email",
  "op": "endsWith",
  "values": [".edu"],
  "negate": false
}

• Multiple clauses = AND (all must match)
• Multiple values = OR (any can match)
• negate: true inverts the operator

Error Handling

Next Steps

After configuring targeting:

1. Provide config URL:

   https://app.launchdarkly.com/projects/{projectKey}/ai-configs/{configKey}

2. Check health with get-agent-configs and section: "health"
3. Attach judges with launchdarkly-online-evals

Related Skills

• launchdarkly-configs-create - Create configs with variations
• launchdarkly-configs-variations - Manage variations
• launchdarkly-configs-update - Restore an earlier version of a variation
• launchdarkly-online-evals - Attach judges

Other Resources

• Target with AgentControl
• Targeting Rules