NVIDIA-NeMo · mmogallapalli · Jul 31, 2026 · Jul 30, 2026 · Jul 30, 2026 · Jul 30, 2026
@@ -236,6 +236,8 @@ nemo-agents-plugin = [
   "pyyaml>=6.0",
   "anthropic>=0.88.0",
   "rich>=13.7.1",
+  "nemo-fabric[claude,codex,deepagents,relay]>=0.1.0rc6,<0.2.0",
+  "nemo-fabric-adapters-hermes>=0.1.0rc6,<0.2.0; python_version < '3.14'",
 ]
 
 # Generated from [tool.bundle-package]; do not edit by hand.

@@ -89,7 +89,7 @@ the platform seed job. For local development, create it manually:
 ```bash
 # Store the API key as a secret
 nemo secrets create ngc-api-key \
-    --data "$NVIDIA_API_KEY"
+    --value "$NVIDIA_API_KEY"
 
 # Create the model provider
 nemo inference providers create nvidia-build \

@@ -1,77 +1,101 @@
 # NeMo Agent Config
 
-This directory contains example Platform-owned `agent.yaml` configs for
+This directory contains Platform-owned `nemo-agents-spec-v1` configs for
 Fabric-backed NeMo Agents. Run the commands below from the repository root.
 
-Fabric dependencies are currently optional so the default workspace does not
-force other Fabric consumers onto the current Fabric SDK/runtime API before they
-migrate.
+Fabric, Relay, Claude, Codex, and DeepAgents dependencies are installed with
+the `nemo-agents` plugin. Hermes is intentionally split out because the Hermes
+Agent runtime dependencies conflict with the Platform environment.
 
-```bash
-uv pip install -e "plugins/nemo-agents[fabric]"
-```
+Shared Fabric capabilities live at the top level:
 
-For Codex/Claude Relay telemetry, install the Relay CLI separately:
+```yaml
+instructions:
+  system:
+    content: You are a concise test assistant.
 
-```bash
-cargo install nemo-relay-cli --root .venv
-.venv/bin/nemo-relay --version
-```
+skills:
+  paths: []
 
-Top-level `skills`, `mcp`, and `tools` are Platform-owned shared fields that
-translate into `FabricConfig`. Prompt settings are harness-specific for now and
-should be configured under `harnesses.<name>.settings`.
+mcp:
+  servers: {}
 
-## Local invoke
+tools:
+  blocked: []
+```
 
-`agent.yaml` is the telemetry-neutral example. Authenticate Codex, then run:
+`instructions.system` is the shared system prompt path for Claude, Codex,
+DeepAgents, and Hermes. Adapter-specific options stay under
+`harnesses.<name>.settings`; do not put prompt text there.
 
-```bash
-codex login
+The selected harness is controlled by `default_harness`. To try another harness
+from the same config today, edit `default_harness` before creating or invoking
+the agent.
 
-nemo agents invoke \
-  --agent-config plugins/nemo-agents/examples/nemo-agent-config/agent.yaml \
-  --input "Reply with exactly: platform fabric works"
-```
+## Invoke
 
-To try Claude, authenticate Claude Code first:
+`agent.yaml` is the telemetry-neutral multi-harness example. Set
+`default_harness` to the harness you want to validate, then create, deploy, and
+invoke the agent through Platform.
 
 ```bash
-claude
-```
+make bootstrap-python
+source .venv/bin/activate
 
-Temporarily set `default_harness: claude` in `agent.yaml`, then run the same
-`nemo agents invoke` command.
+export NVIDIA_API_KEY="<your NVIDIA API key>"
 
-## Relay Local Files
+nemo setup --auto --start-services --install-skills --no-deploy-agent
 
-`agent-relay.yaml` enables Relay telemetry without Intake. It writes local ATIF
-and ATOF artifacts under the deployment artifacts directory.
+# If setup does not create a usable NVIDIA inference provider, follow
+# Step 2 in plugins/nemo-agents/README.md before deploying.
 
-```bash
 nemo agents create \
-  --name fabric-relay-local-test \
-  --agent-config plugins/nemo-agents/examples/nemo-agent-config/agent-relay.yaml
+  --name fabric-agent \
+  --agent-config plugins/nemo-agents/examples/nemo-agent-config/agent.yaml
 
 nemo agents deploy \
-  --agent fabric-relay-local-test \
-  --name fabric-relay-local-test-deployment \
+  --agent fabric-agent \
+  --name fabric-agent-deployment \
   --mode subprocess
 
 nemo agents invoke \
-  --agent-deployment fabric-relay-local-test-deployment \
-  --input "Reply with exactly: relay local works"
+  --agent-deployment fabric-agent-deployment \
+  --input "Reply with exactly: platform fabric works"
 ```
 
-Confirm Relay emitted both ATIF and ATOF files:
+Use a unique `--name` / deployment name for each harness, or delete the previous
+agent and deployment before recreating them.
+
+## Harness Notes
+
+### Codex
+
+Set `default_harness: codex` in `agent.yaml`. Authenticate Codex before
+invoking:
 
 ```bash
-find ~/.local/share/nemo/agents/system/default/fabric-relay-local-test-deployment-fabric/artifacts \
-  \( -name "*atif*" -o -name "*atof*" \) \
-  -exec ls -lh {} \;
+codex login
+```
+
+In this example, Codex uses the shared Nemotron model through Platform IGW.
+
+### DeepAgents
+
+Set `default_harness: deepagents` in `agent.yaml`. In this example, DeepAgents
+uses the shared Nemotron model through Platform IGW.
+
+### Claude
+
+Set `default_harness: claude` in `agent.yaml`. Authenticate Claude Code before
+invoking:
+
+```bash
+claude
 ```
 
-## Hermes
+In this example, Claude uses its harness-local Anthropic model config.
+
+### Hermes
 
 Hermes Agent has dependencies that conflict with the Platform environment, so
 install it with the Fabric adapter in a separate Python 3.12 environment:
@@ -80,43 +104,40 @@ install it with the Fabric adapter in a separate Python 3.12 environment:
 uvx uv@0.9.14 venv --python 3.12 .venv-hermes
 uvx uv@0.9.14 --no-config pip install \
   --python .venv-hermes/bin/python \
-  "nemo-fabric[relay]>=0.1.0rc4,<0.2.0" \
-  "nemo-fabric-adapters-hermes>=0.1.0rc4,<0.2.0" \
+  "nemo-fabric[relay]>=0.1.0rc6,<0.2.0" \
+  "nemo-fabric-adapters-hermes>=0.1.0rc6,<0.2.0" \
   "hermes-agent==0.19.0"
 
-export HERMES_ADAPTER_PYTHON="$PWD/.venv-hermes/bin/python"
-export NVIDIA_API_KEY="<your NVIDIA API key>"
+export ADAPTER_PYTHON="$PWD/.venv-hermes/bin/python"
 ```
 
-Temporarily set `default_harness: hermes` in `agent.yaml`, then run:
+Set `default_harness: hermes` in `agent.yaml`. For subprocess deployments,
+export `ADAPTER_PYTHON` before starting Platform, or restart Platform after
+exporting it. The Platform service launches the agent subprocess, so exporting
+`ADAPTER_PYTHON` only in the later CLI shell is not enough.
+
+## Relay Local Files
+
+`agent-relay.yaml` enables Relay telemetry without Intake. It writes local ATIF
+and ATOF artifacts under the deployment artifacts directory. Use
+`agent-relay.yaml` with the invoke directions above.
+
+Confirm Relay emitted both ATIF and ATOF files:
 
 ```bash
-nemo agents invoke \
-  --agent-config plugins/nemo-agents/examples/nemo-agent-config/agent.yaml \
-  --input "Reply with exactly: platform hermes works"
+find ~/.local/share/nemo/agents/system/default/fabric-relay-local-test-deployment-fabric/artifacts \
+  \( -name "*atif*" -o -name "*atof*" \) \
+  -exec ls -lh {} \;
 ```
 
-## Relay to Intake (Local)
+## Relay to Intake
 
 `agent-relay-intake.yaml` enables Relay ATIF export to a locally running
-Platform Intake API. Start ClickHouse for Intake, start Platform, then create
-and deploy the agent:
+Platform Intake API. Start ClickHouse for Intake, then use
+`agent-relay-intake.yaml` with the invoke directions above.
 
 ```bash
 services/intake/scripts/spans/run_clickhouse.sh
-
-nemo agents create \
-  --name fabric-relay-intake \
-  --agent-config plugins/nemo-agents/examples/nemo-agent-config/agent-relay-intake.yaml
-
-nemo agents deploy \
-  --agent fabric-relay-intake \
-  --name fabric-relay-intake-deployment \
-  --mode subprocess
-
-nemo agents invoke \
-  --agent-deployment fabric-relay-intake-deployment \
-  --input "Reply with exactly: platform fabric works"
 ```
 
 Confirm Intake received ATIF-derived spans:

@@ -2,37 +2,47 @@ config_format: nemo-agents-spec-v1
 name: test-agent
 description: Test agent config
 
+instructions:
+  system:
+    content: You are a concise test assistant.
+
 default_harness: codex
 
 harnesses:
+  claude:
+    kind: claude
+    model:
+      provider: anthropic
+      model: anthropic/claude-sonnet-4-5
+    settings:
+      permission_mode: dontAsk
   hermes:
     kind: hermes
     model:
       provider: nvidia
       model: nvidia/nemotron-3-nano-30b-a3b
       api_key_env: NVIDIA_API_KEY
+      base_url: https://integrate.api.nvidia.com/v1
       temperature: 0.0
     settings:
-      python_env: HERMES_ADAPTER_PYTHON
-      base_url: https://integrate.api.nvidia.com/v1
-      max_iterations: 1
       max_tokens: 512
       reasoning_config:
         effort: none
-      enabled_toolsets: []
-      system_prompt: You are a concise test assistant.
+  deepagents:
+    kind: deepagents
+    settings:
+      deepagents: {}
   codex:
     kind: codex
     settings:
       sandbox: workspace-write
-      system_prompt: You are a concise test assistant.
-      config_overrides:
-        model_reasoning_effort: high
+      reasoning_effort: high
 
 models:
   default:
-    provider: openai
-    model: openai/gpt-5.4
+    provider: nvidia
+    model: nvidia-nemotron-3-nano-30b-a3b
+    api_key_env: NVIDIA_API_KEY
 
 skills:
   paths: []

@@ -2,37 +2,47 @@ config_format: nemo-agents-spec-v1
 name: test-agent
 description: Test agent config
 
+instructions:
+  system:
+    content: You are a concise test assistant.
+
 default_harness: codex
 
 harnesses:
+  claude:
+    kind: claude
+    model:
+      provider: anthropic
+      model: anthropic/claude-sonnet-4-5
+    settings:
+      permission_mode: dontAsk
   hermes:
     kind: hermes
     model:
       provider: nvidia
       model: nvidia/nemotron-3-nano-30b-a3b
       api_key_env: NVIDIA_API_KEY
+      base_url: https://integrate.api.nvidia.com/v1
       temperature: 0.0
     settings:
-      python_env: HERMES_ADAPTER_PYTHON
-      base_url: https://integrate.api.nvidia.com/v1
-      max_iterations: 1
       max_tokens: 512
       reasoning_config:
         effort: none
-      enabled_toolsets: []
-      system_prompt: You are a concise test assistant.
+  deepagents:
+    kind: deepagents
+    settings:
+      deepagents: {}
   codex:
     kind: codex
     settings:
       sandbox: workspace-write
-      system_prompt: You are a concise test assistant.
-      config_overrides:
-        model_reasoning_effort: high
+      reasoning_effort: high
 
 models:
   default:
-    provider: openai
-    model: openai/gpt-5.4
+    provider: nvidia
+    model: nvidia-nemotron-3-nano-30b-a3b
+    api_key_env: NVIDIA_API_KEY
 
 skills:
   paths: []

@@ -2,6 +2,10 @@ config_format: nemo-agents-spec-v1
 name: test-agent
 description: Test agent config
 
+instructions:
+  system:
+    content: You are a concise test assistant.
+
 default_harness: codex
 
 harnesses:
@@ -18,28 +22,27 @@ harnesses:
       provider: nvidia
       model: nvidia/nemotron-3-nano-30b-a3b
       api_key_env: NVIDIA_API_KEY
+      base_url: https://integrate.api.nvidia.com/v1
       temperature: 0.0
     settings:
-      python_env: HERMES_ADAPTER_PYTHON
-      base_url: https://integrate.api.nvidia.com/v1
-      max_iterations: 1
       max_tokens: 512
       reasoning_config:
         effort: none
-      enabled_toolsets: []
-      system_prompt: You are a concise test assistant.
+  deepagents:
+    kind: deepagents
+    settings:
+      deepagents: {}
   codex:
     kind: codex
     settings:
       sandbox: workspace-write
-      system_prompt: You are a concise test assistant.
-      config_overrides:
-        model_reasoning_effort: high
+      reasoning_effort: high
 
 models:
   default:
-    provider: openai
-    model: openai/gpt-5.4
+    provider: nvidia
+    model: nvidia-nemotron-3-nano-30b-a3b
+    api_key_env: NVIDIA_API_KEY
 
 skills:
   paths: []