Agent Integration Catalog
Back to main doc
This directory contains one document per agent type/framework describing its capabilities, Rossoctl integration points, and testing status.
Agent Classification
| Category | Description | Examples |
|---|---|---|
| Custom A2A | Agents built with specific SDKs, deployed as K8s Deployments, accessed via A2A JSON-RPC | weather, ADK, Claude SDK |
| OpenShell Builtin | Pre-installed CLI agents in the OpenShell base sandbox image, accessed via SSH/exec | Claude Code, OpenCode, Codex, Copilot |
| Future | Agents and frameworks planned for integration | LangGraph, CrewAI, AutoGen, Cursor |
Current Agents
Deployed and Tested
| Agent Doc | Type | Framework | LLM | Supervisor | Status |
|---|---|---|---|---|---|
| claude-sdk-agent.md | Custom A2A | Anthropic SDK | LiteMaaS | No | All skills pass |
| adk-agent.md | Custom A2A | Google ADK | LiteMaaS | Yes (supervised) | All skills pass |
| weather-supervised.md | Custom A2A | LangGraph | No | Yes (supervised) | Infra tests pass |
| openshell-claude.md | Builtin | Claude Code CLI | Anthropic/LiteLLM | Yes | Sandbox CR tested |
| openshell-opencode.md | Builtin | OpenCode CLI | OpenAI-compat | Yes | All LiteMaaS skills pass |
NemoClaw Agents (integrated from NVIDIA/NemoClaw)
| Agent | Type | Framework | LLM | Status |
|---|---|---|---|---|
| nemoclaw-hermes | NemoClaw | Nous Research Hermes (v2026.4.13) | LiteMaaS | Platform + security tests pass |
| nemoclaw-openclaw | NemoClaw | OpenClaw (v2026.4.9) | LiteMaaS | All tests pass |
Future agents (planned)
| Agent | Type | Framework | Notes |
|---|---|---|---|
| openshell_codex | Builtin | OpenAI Codex | Needs real OpenAI key |
| openshell_copilot | Builtin | GitHub Copilot | Needs GitHub subscription |
| langgraph_complex | Custom A2A | LangGraph (multi-node) | Multi-node graph agents |
| crewai_crew | Custom A2A | CrewAI | Multi-agent crew orchestration |
Standard Document Format
Every agent doc MUST follow this structure. This ensures consistency and makes the capability matrix scannable across agents.
# {Agent Name}
> **Type:** Custom A2A | Builtin Sandbox
> **Framework:** {framework name}
> **LLM:** {provider} | None
> **Supervisor:** Yes | No
> **Status:** Deployed | Planned | Blocked
## 1. Overview
One paragraph: what the agent does and why it exists.
## 2. Architecture
Mermaid diagram showing: agent pod → LLM → tools/APIs.
## 3. Files
List of deployment files with paths.
## 4. Deployment
Commands to build, deploy, and configure on Kind and OCP.
## 5. Capabilities
Table of agent capabilities with status:
| Capability | Supported | Notes |
|-----------|-----------|-------|
| A2A protocol | Yes/No | |
| Multi-turn context | Yes/No/Partial | How context is managed |
| Tool calling | Yes/No | Which tools |
| Subagent delegation | Yes/No | |
| Memory/knowledge | Yes/No | What persists |
| Skill execution | Native/Prompt/No | How rossoctl skills are loaded |
| HITL approval | L0-L3 | Which HITL level |
## 6. Rossoctl Integration
How this agent connects to Rossoctl platform services:
### 6.1 Communication Adapter
How the Rossoctl backend talks to this agent:
- A2A JSON-RPC (custom agents)
- ExecSandbox gRPC (builtin sandboxes)
- kubectl exec (fallback)
### 6.2 Session Management
How conversation history is stored and resumed:
- Backend PostgreSQL (A2A agents)
- Workspace PVC (builtin sandboxes)
- Agent-side storage (if any)
### 6.3 Observable Events
What data the agent produces that Rossoctl can display:
| Event | Source | Rossoctl UI Component | Phase |
|-------|--------|---------------------|-------|
| LLM request/response | Agent SDK | PromptInspector | Phase 2 |
| Tool call result | Agent framework | EventsPanel | Phase 2 |
| Subagent dispatch | Agent framework | SubSessionsPanel | Phase 3 |
| File modification | Workspace PVC | FileBrowser | Phase 2 |
| Budget usage | LiteLLM/Budget Proxy | LlmUsagePanel | Phase 2 |
| HITL approval request | Supervisor OPA | HitlApprovalCard | Phase 3 |
### 6.4 FileBrowser Integration
What workspace files the agent creates that FileBrowser can display:
| Path | Content | Browsable |
|------|---------|-----------|
| /workspace/.claude/ | Session transcripts | Yes |
| /workspace/project/ | Agent-modified code | Yes |
## 7. LLM Compatibility
Which LLM providers work with this agent:
| Provider | Protocol | Works? | Notes |
|----------|----------|--------|-------|
| LiteMaaS | OpenAI-compat | Yes/No | |
| Anthropic API | Claude messages | Yes/No | |
| Ollama | OpenAI-compat | Yes/No | |
## 8. Policy Configuration
OPA/Rego policy for this agent:
| Policy | Value | Notes |
|--------|-------|-------|
| Filesystem read-only | /usr, /etc | |
| Filesystem read-write | /tmp, /app | |
| Network egress | *.svc.cluster.local | |
## 9. Testing Status
Which E2E tests cover this agent:
| Test File | Tests | Pass | Skip | Fail |
|-----------|-------|------|------|------|
| test_02_a2a_connectivity | 2 | 2 | 0 | 0 |
| test_07_skill_execution | 3 | 1 | 2 | 0 |
## Capability Comparison Matrix
Quick cross-agent comparison of key capabilities:
| Capability | weather | adk | claude_sdk | supervised | openshell_claude | openshell_opencode |
|-----------|---------|-----|------------|------------|-----------------|-------------------|
| **A2A protocol** | Yes | Yes | Yes | Yes (exec) | No (SSH/exec) | No (SSH/exec) |
| **LLM calls** | No | Yes | Yes | No | Yes | Yes |
| **Tool calling** | MCP | ADK tools | Prompt | MCP | Native | Native |
| **Multi-turn** | Stateless | contextId | Stateless | Stateless | Terminal session | Terminal session |
| **Subagents** | No | No | No | No | Yes (native) | No |
| **Memory** | No | In-memory | No | No | Disk (.claude/) | Disk (.opencode/) |
| **Skill execution** | No | Via prompt | Via prompt | No | Native (.claude/skills/) | Via prompt |
| **HITL** | N/A | L0 | L0 | L0 (OPA) | L0-L3 | L0-L3 |
| **Workspace PVC** | No | No | No | No | Yes | Yes |
| **FileBrowser** | N/A | N/A | N/A | N/A | .claude/ sessions | .opencode/ sessions |
| **OTel traces** | No | Partial | Partial | No | No | No |
| **Budget tracking** | N/A | LiteLLM | LiteLLM | N/A | Gateway | Gateway |
## A2A Adapter Architecture
For agents that don't natively speak A2A, Rossoctl needs an adapter:
```mermaid
graph TB
subgraph backend["Rossoctl Backend"]
A2A["A2A Adapter"]
EXEC["ExecSandbox Adapter"]
TERM["Terminal Adapter"]
end
subgraph agents["Custom A2A"]
W["weather_agent"]
ADK["adk_agent"]
CS["claude_sdk_agent"]
end
subgraph sandboxes["Builtin Sandboxes"]
GW["OpenShell Gateway"]
CC["Claude Code"]
OC["OpenCode"]
end
A2A -->|"message/send"| W & ADK & CS
EXEC -->|"ExecSandbox gRPC"| GW
GW -->|"exec in pod"| CC & OC
TERM -->|"WebSocket → SSH"| GW
| Agent Type | Adapter | Protocol | Implemented? |
|---|---|---|---|
| Custom A2A | A2A Adapter | JSON-RPC over HTTP | Yes (current) |
| Builtin sandbox | ExecSandbox Adapter | gRPC streaming | Phase 2 |
| Interactive sandbox | Terminal Adapter | WebSocket → SSH | Phase 3 |