docs: define Resources/Agent/Model Server in the glossary

Add Resources Server / Agent Server / Model Server entries to the Architecture Terms glossary.
Per review: do not equate the resources server with "the environment" (in Gym an environment is a
resources server + agent + dataset(s), see environments/), and surface multiple built-in agent
harnesses (simple_agent, aviary_agent, ...) rather than featuring simple_agent alone.

Signed-off-by: Wojciech Prazuch <wprazuch@nvidia.com>

Author: wprazuch

fern/versions/latest/pages/about/concepts/key-terminology.mdx

@@ -57,6 +57,18 @@ Training data format for DPO consisting of the same prompt with two different re
 
 The primary LLM being trained or evaluated - the "decision-making brain" you want to improve.
 
+**Resources Server**
+
+The FastAPI service (under `resources_servers/`) that holds per-task state, exposes the environment's tools, and implements `verify()` to score a rollout into a reward — the verifier-and-state component of an environment. (In Gym an *environment* is a resources server together with an agent and dataset(s); see the `environments/` directory.) Every environment has exactly one resources server.
+
+**Agent Server (Responses API Agent)**
+
+The FastAPI service (under `responses_api_agents/`) that drives the model through a task — the harness that runs the multi-step / tool-calling loop against a resources server. Gym ships several built-in harnesses (e.g. `simple_agent`, `aviary_agent`, and others under `responses_api_agents/`); pick whichever fits your control flow, or bring your own.
+
+**Model Server (Responses API Model)**
+
+The FastAPI service (under `responses_api_models/`) that wraps an LLM endpoint (vLLM, OpenAI, etc.) behind NeMo Gym's Responses API and manages the token IDs needed for training. Agents refer to it by name (commonly `policy_model`).
+
 **Orchestration**
 
 Coordination logic that manages when to call models, which tools to use, and how to sequence multi-step operations.

fern/versions/v0.3.0/pages/about/concepts/key-terminology.mdx

@@ -57,6 +57,18 @@ Training data format for DPO consisting of the same prompt with two different re
 
 The primary LLM being trained or evaluated - the "decision-making brain" you want to improve.
 
+**Resources Server**
+
+The FastAPI service (under `resources_servers/`) that holds per-task state, exposes the environment's tools, and implements `verify()` to score a rollout into a reward — the verifier-and-state component of an environment. (In Gym an *environment* is a resources server together with an agent and dataset(s); see the `environments/` directory.) Every environment has exactly one resources server.
+
+**Agent Server (Responses API Agent)**
+
+The FastAPI service (under `responses_api_agents/`) that drives the model through a task — the harness that runs the multi-step / tool-calling loop against a resources server. Gym ships several built-in harnesses (e.g. `simple_agent`, `aviary_agent`, and others under `responses_api_agents/`); pick whichever fits your control flow, or bring your own.
+
+**Model Server (Responses API Model)**
+
+The FastAPI service (under `responses_api_models/`) that wraps an LLM endpoint (vLLM, OpenAI, etc.) behind NeMo Gym's Responses API and manages the token IDs needed for training. Agents refer to it by name (commonly `policy_model`).
+
 **Orchestration**
 
 Coordination logic that manages when to call models, which tools to use, and how to sequence multi-step operations.