Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
A tuning environment in Frontier Tuning is a container for the resources an AI system uses to attempt business tasks and assess its responses. Those resources include skills, evaluation criteria, sample prompts, knowledge sources, tools, model settings, and execution history.
As an agent creator, you use these building blocks to describe a task, define a good response, and compare changes. This article explains the concepts shared across tuning workflows, rather than the steps for a particular task or plugin.
Tuning environments
A tuning environment, also called a reinforcement learning environment (RLE), brings the resources for a tuning scenario together. The resources have different purposes:
| Resource | Purpose |
|---|---|
| Skills | Describe how the agent responds to a type of request. |
| Evaluation criteria | Define the standards for assessing a skill's responses. |
| Sample prompts | Provide example requests for evaluation or training. |
| Knowledge sources | Ground responses in your content. |
| Connectors and tools | Expose actions the agent can take. |
| Model settings | Identify models available for responses, evaluation, or fine-tuning. |
| Execution history | Record runs through the environment. |
In CLI commands, a tuning environment is called an environment, as in frontier-tuning environments create. The -e option identifies the environment a command acts on.
Skills and request routing
A skill defines how the agent responds to one type of request. It combines:
- Instructions that the agent follows.
- A description used to route requests to the skill.
- Evaluation criteria used to assess its responses.
- Optional knowledge references for that skill.
The description helps the environment select a matching skill. The instructions describe how to handle the request after the skill is selected. Skills can be enabled or disabled for routing.
Where skill versioning is available, version history lets you inspect earlier versions. Versioning availability depends on your tenant.
Evaluation criteria
Evaluation criteria are the standards used to assess a skill's responses. A criterion describes what a good response includes, such as citing a policy or stating a deadline.
In CLI commands and skill files, evaluation criteria are called rubrics, as in frontier-tuning rubrics list. A criterion includes:
- A name.
- One or more checklist items.
- An importance level:
critical,high,medium, orlow. - An enabled state.
Criteria connect your expectations to response grading. Depending on the operation, they can also guide answer selection or provide a signal for fine-tuning. These uses are distinct from the instructions that tell the agent how to respond.
Sample prompts and held-out evaluation
A sample prompt is an example request used for evaluation or fine-tuning. A sample belongs to a skill and can reference a file, such as a SharePoint document that the request asks the agent to summarize.
| Sample label | Purpose |
|---|---|
| Evaluation | Assess responses against evaluation criteria. |
| Training | Provide examples for fine-tuning or for refining skill instructions and evaluation criteria. |
The label describes the sample's intended use. Don't rely on the label alone to keep evaluation data separate from improvement inputs.
A held-out set contains Evaluation samples that you don't use for generation, refinement, or training. Keeping those requests separate helps you assess responses to examples that weren't used to make the improvement.
Samples are graded with their owning skill's evaluation criteria. A change to those criteria can change the meaning of the resulting scores, even when the prompts stay the same.
Knowledge sources
A knowledge source grounds the agent in your content. Sources can include a SharePoint site or file, a OneDrive or SharePoint folder, a Teams thread, or a meeting.
Knowledge references identify content for the agent to use. They're distinct from skill instructions, which describe how to respond, and from tools, which expose actions such as searching for or reading content.
Connectors and tools
A connector groups tools for a service, such as Outlook Mail or SharePoint. A tool is one action exposed through that connector, such as searching mail or reading a file.
Connector configuration and individual tool enablement are different controls. Some connector types support individual tool toggles; the available controls depend on the connector type.
Tools also differ in the services and data they use:
| Tool type | Intended behavior |
|---|---|
| Real tool | Calls the actual service and uses your data. |
| Virtualized tool | Uses the real service for reads and intercepts supported writes during evaluation and training. |
| Mock tool | Uses deterministic, canned data as a stand-in for a real tool. |
These distinctions help you understand what a run interacts with. The type name alone doesn't establish that every action is intercepted or that a run can't affect a service.
Connectors can have stage labels such as Inference, Evaluation, and Finetuning, with Tuning shown in CLI tables. Stage selection is separate from the tool's real, virtualized, or mock behavior.
Models and inference strategies
A model and an inference strategy describe different parts of a response run. Model settings identify the model used. An inference strategy, also called a harness, controls how an execution searches for an answer.
The CLI groups available models by purpose:
| Model list | Purpose |
|---|---|
| Models for Evaluation | Models available for evaluation with evaluate start, including models that might be evaluation-only. |
| Models for Inferencing | Models available for skill responses, including fine-tuned models, and their enabled state. |
| Models for Fine-tuning | Base models available for fine-tuning with tune start. |
The enabled inferencing models apply to skill responses. Don't assume that every model in an evaluation list is also available for responses or fine-tuning.
Inference strategies vary in how they generate and assess candidate responses:
| Strategy | Approach |
|---|---|
Simple |
Runs a single execution. |
BestOfN |
Generates several candidates and selects the best-graded result. |
TreeSearch |
Explores candidate paths guided by evaluation criteria. |
SimpleNoEvaluation |
Runs a single response without grading. |
Available strategies depend on the operation. For example, a strategy available for an execution isn't necessarily available for an evaluation job.
Executions, conversations, and improvement jobs
An execution is one run of a query through a tuning environment. Executions that share a conversation ID form a multi-turn conversation.
Evaluation, refinement, and fine-tuning have different outcomes:
| Operation | Outcome |
|---|---|
| Evaluation job | Runs samples and grades the responses against evaluation criteria. |
| Refinement run | Proposes changes to skill instructions, evaluation criteria, or both, using Training samples. |
| Fine-tuning job | Trains a tuned model from a base model. |
| Sample generation run | Produces candidate sample prompts for review. |
Applying a refinement changes skill artifacts; fine-tuning produces a model. An evaluation measures responses rather than making either kind of change.
Review and save behavior depends on the operation. A refinement proposal and a generation operation don't necessarily use the same approval or persistence steps.
Baselines and response grading
A baseline is an evaluation you run before a change. Comparing it with an evaluation after the change helps you assess the result.
Keep the samples, evaluation criteria, model, and strategy the same for the comparison. If these inputs change, the score difference doesn't describe only the improvement you're assessing.
Trajectory grading assesses an agent's tool calls and intermediate steps, rather than only its final response. It describes a different grading scope, not a separate name for every evaluation.
Participants, access, and governance
A maker builds and improves a tuning environment. A subject-matter expert (SME) contributes the expertise that defines tasks, procedures, and evaluation criteria. These descriptions explain people's contributions; they aren't access assignments.
Environment roles describe access within Frontier Tuning. The role names include Owner, Contributor, and User. They're distinct from Azure role-based access control (RBAC) assignments and Microsoft Entra directory roles.
Governance and licensing are separate from the concepts used to describe and assess a task. For the current discussion of SharePoint Embedded containers, Microsoft Agent 365, access, and cost categories, see the Frontier Tuning admin guide.