LaunchDarkly AI

LaunchDarkly AI is LaunchDarkly’s AI-focused platform for helping development teams safely build, test, deploy, monitor, and control AI-powered applications and agents in production. Its current AI offering is centered on AgentControl, which extends LaunchDarkly’s established feature-management and release-control capabilities to AI agents and large language model (LLM) applications.

LaunchDarkly was founded in 2014 by Edith Harbaugh and John Kodumal. The company originally focused on feature management and progressive software releases. Today, LaunchDarkly positions itself as a runtime control layer for modern software and AI development.

AgentControl allows teams to keep prompts, model settings, instructions, and other AI configuration outside their application code. This means developers can modify AI behavior, test variations, change models, and progressively release updates without having to redeploy the application each time.

A key benefit is the ability to manage AI behavior throughout the development lifecycle. Teams can create different prompt and model variations, use targeting rules to determine which configuration users receive, compare variations, and evaluate AI behavior using metrics such as quality, cost, latency, and satisfaction.

AI Agent Management

AgentControl supports both completion mode and agent mode. Completion mode is designed for single-step model responses using messages and roles, while agent mode supports structured, multi-step workflows. External tools and APIs can be incorporated into either mode depending on the application and SDK implementation.

Teams can also use AgentControl to progressively introduce new AI models or configurations. Instead of changing an AI system for every user simultaneously, organizations can target specific audiences and gradually increase traffic to a new configuration.

AI Evaluation and Experimentation

LaunchDarkly AI provides tools for evaluating AI behavior before and after deployment. Teams can use offline evaluations to compare configurations against datasets and online evaluations to continuously assess production behavior. LLM-based judges can be used to score responses according to quality signals such as accuracy, relevance, toxicity, or custom criteria.

The platform also supports A/B testing and multi-armed bandit experimentation. This allows teams to compare prompts, models, or agent configurations using real-world traffic and business metrics rather than relying solely on development-time testing.

AI Observability and Runtime Control

LaunchDarkly AI is particularly focused on what happens after an AI application reaches production. AgentControl provides monitoring for AI model generations, including metrics such as generation counts, token usage, latency, errors, and evaluation scores. These metrics can be aggregated and displayed through LaunchDarkly’s monitoring capabilities.

The platform also supports adaptive runtime controls. For example, teams can define thresholds and configure their systems to respond when quality or other performance signals change. LaunchDarkly describes AgentControl as enabling agents to respond to changing production conditions without requiring a traditional redeployment cycle.

Model Flexibility

LaunchDarkly does not act as the AI model provider or proxy model requests on behalf of applications. Instead, AgentControl supplies model configuration information to the application, while the application remains responsible for calling the selected AI provider. This approach allows organizations to work with supported providers and frameworks while retaining control over configuration and rollout.

The platform supports standard models available through LaunchDarkly as well as custom AI model configurations. Teams can therefore configure their own models and parameters when required.

LaunchDarkly also provides dedicated AI SDKs for .NET, Go, Node.js, Python, and Ruby. The AI SDKs are currently pre-1.0, with Python and Node.js AI SDKs described by LaunchDarkly as open beta/high-level SDKs for working with models, judges, and agent graphs.

Developer and Enterprise Focus

LaunchDarkly AI is primarily aimed at engineering, AI, DevOps, SRE, and product teams rather than being a general-purpose consumer chatbot. Its strongest value is in organizations that need governance, experimentation, observability, controlled rollouts, and operational safeguards around AI applications.

The wider LaunchDarkly platform combines AI capabilities with feature flags, progressive rollouts, automated rollbacks, observability, experimentation, and integrations with development and operational tools. This creates a unified environment for controlling both traditional software releases and AI behavior.

For organizations adopting AI at scale, LaunchDarkly AI is most useful when the challenge is not simply generating AI responses but controlling how AI behaves in production. Its combination of configuration management, evaluations, experimentation, monitoring, targeting, and runtime control makes it particularly relevant for teams building customer-facing AI applications and autonomous or semi-autonomous agents.

Pricing

LaunchDarkly currently uses a usage-based pricing model rather than relying primarily on per-seat pricing.

Developer

• $0/month
• Free forever
• Unlimited seats
• Unlimited feature flags
• 100K experimentation MAU/month included
• 5,000 AI runs/month included
• 14-day AI data retention
• Limited service connections and observability usage

Foundation

• Pay-as-you-go
• 14-day trial
• No platform or seat fees
• $10 per service connection/month when billed yearly
• $8.33 per 1,000 client-side MAU/month when billed yearly
• 5,000 AI runs/month included
• Additional AI runs: $5 per 1,000
• 30-day AI data retention

Enterprise

• Custom pricing
• Designed for organizations requiring advanced controls and enterprise-scale capabilities
• Custom AI-run allowances and longer data retention
• Enterprise security and governance capabilities

AgentControl is an add-on feature, and availability depends on the organization’s LaunchDarkly plan.

Pricing Note: LaunchDarkly’s pricing is primarily usage-based. Costs can depend on service connections, monthly active users, AI runs, experimentation, observability, and other usage metrics. AI runs beyond included allowances can generate additional charges on applicable plans.

Review

LaunchDarkly AI is a strong choice for development teams that need more than a basic AI API or chatbot. Its main advantage is the ability to treat AI behavior as something that can be configured, evaluated, experimented with, monitored, and safely changed in production.

AgentControl is particularly compelling for teams operating AI agents because prompts, model configurations, and instructions can be managed separately from application code. This reduces the need for redeployments when teams want to test or modify AI behavior. Its targeting, evaluation, experimentation, observability, and runtime-control capabilities also provide a structured approach to operating AI systems after deployment.

The main limitation is that LaunchDarkly AI is not designed to replace an LLM provider. Applications still need to call their chosen model provider themselves, while LaunchDarkly manages configuration, evaluation, rollout, and monitoring. AgentControl is also an add-on whose availability depends on the customer’s plan.

For organizations already using LaunchDarkly, its AI capabilities can be a natural extension of existing release-management workflows. For teams primarily looking for a simple standalone AI assistant or model API, however, LaunchDarkly AI may provide more infrastructure and governance functionality than they need.