AI Configs

Change AI behavior instantly, no redeploy needed

Tune prompts, swap models, and adjust agent parameters in real time. Governed by the same approval workflows and audit trails that protect your production deployments.

Why AI Configs

Everything you need to run AI safely at runtime.

Define the schema, tune the values, and prove every change works, all without a deployment.

Define

Set a structured schema once. Dev, staging, and production each manage their own values without touching code.

Tune

Update prompts, models, and parameters at runtime, no deployment required, with full approval and audit trails.

Experiment

Measure quality, cost, and latency before you promote. Ship the winning config with confidence.

Define

Set the schema once for every environment.

AI Configs use a structured schema to define the expected shape of your prompts, models, and parameters. Set the structure once at the config level, dev, staging, and production each manage their own values independently.

AI config schemas. Start from a structured template, or import a prompt directly from Harness AI Evals. The schema defines what's valid; your team fills in the values without touching code.

Per-environment values. Dev, staging, and production each hold their own config values against the same schema. Iterate fast in staging without ever touching production.

One-click value promotion. When a config is ready, promote its values across environments in one governed action. Approval required, audit trail automatic. Not copy-paste.

Tune

Update any AI parameter at runtime.

Prompts, models, temperature, update any AI parameter without a deployment. Every change goes through the same approval workflows and audit trails that protect your feature releases.

Swap models in real time. Move from GPT-4o to Claude or a fine-tuned model without writing code. Changes propagate in seconds, not sprints.

Governed prompt updates. Every config change goes through the same approval workflow as your feature releases. Full audit trail, required reviews, and rollback on demand.

Serve different variations to different users. Define multiple config variations and target them by user segment, plan tier, region, or any attribute. Same config object, different values per audience.

Experiment

Measure quality, cost, and latency before you promote.

Don't guess which model or prompt is better. Run controlled experiments on your AI configs against real traffic. Measure what matters, output quality, token cost, and latency, then promote the winner with confidence.

Quality vs. cost tradeoffs. Measure output quality scores alongside token cost and latency. Find the right model for each use case, not just the most capable one.

Prompt iteration at speed. Test prompt variants against each other on live traffic. No staging environment, no synthetic data. Real users, real results.

Promote with confidence. When an experiment reaches statistical significance, promote the winning config to all users in one click. Full history preserved.

What our customers say

Teams are releasing features faster with Harness

10x

faster flag setup

<3ms

cached response time

Before Harness it could be anywhere from 1 - 3 weeks to get your flag up and running. Now we're able to do it in 1 to 3 days ... We're getting cached responses in under three milliseconds.

Patrick Laughlin, Senior Software Engineer, ADP

Read more
FAQ

Frequently asked questions

AI config management is the practice of storing, versioning, and delivering the dynamic inputs that control AI behavior at runtime. This includes prompts, model parameters, temperature settings, system instructions, and tool definitions. Instead of hardcoding these values in your application, an AI config management platform like Harness lets you update them instantly, target them to specific user segments, and roll back when something goes wrong, all without a code deployment.

Harness AI Configs stores prompt templates and model parameters as versioned, JSON-based configurations in a central platform. Your application fetches the active config at runtime through the Harness SDK. When you need to change a prompt, adjust a model, or tune a parameter, you update the config in Harness and the change propagates instantly to live traffic. No code change, no pull request, no deployment pipeline.

Yes. Any AI config in Harness can be turned into a multi-variant experiment. You define two or more config variations, such as different prompts, different model versions, or different temperature settings, and Harness splits traffic between them using the same bucketing engine as feature flags. You connect a quality or cost metric, and Harness tracks which variation performs better so you can ship the winning config with confidence.

Harness applies the same governance layer to AI configs that it applies to feature flags. Every config change goes through role-based access control (RBAC), optional OPA policy checks, and configurable approval workflows before it reaches production. A full audit trail records who changed what and when. This means your AI behavior is version-controlled and auditable, not ad-hoc and invisible.

Harness AI Configs is provider-agnostic. You can manage configurations for any LLM provider, including OpenAI, Anthropic, Google, Mistral, and self-hosted models. The platform stores the config values; your application code handles the actual API call. This gives you the flexibility to switch providers or models through a config update rather than a code change.

A database gives you storage but none of the runtime delivery features you need in production. Harness adds per-user and per-environment targeting, percentage rollouts, instant rollback, experiment assignment, approval workflows, and a real-time SDK with sub-200ms evaluation latency. It also enforces a JSON-based schema so every config change is validated before it ships, preventing malformed values from reaching production.

Get started with Harness AI Configs

Try it free. Update models, tune prompts, and run experiments on live traffic, no deployment required.