1What "agent" means here
Not autonomous software making decisions for us. Here an agent is the market's working definition: an AI-powered application that executes a specific function against our grounded data. The platform offers a portfolio of them, plus two ways for other software to reach the same knowledge:
MCP server
Model Context Protocol. Lets AI assistants we already own call Trilivy as a governed tool, and lets Trilivy appear as a connector inside those platforms.
REST API
258 endpoints for legacy systems, intranet pages, scheduled jobs and third-party software. No MCP needed.
Tool-using agent
One "ask" endpoint. The model selects from registered tools (SQL, REST, retrieval, Python, existing platform pages) to answer a question.
2The portfolio
| Agent | Function |
|---|---|
| Action Registry & agent tools | 39 registered actions and 13 tools the chatbot can invoke; hot-reload, no restart |
| Content Studio | Author-Revise-Format pipeline that writes finished documents from our corpus (C5) |
| RepGen | Business-intelligence reports and operations playbooks, with history and downloads |
| Persona Splitter | Reframes one document for executive, technical and adopter audiences |
| Web Crawler | Turns a public site or intranet into queryable knowledge |
| Course Maker & AI Learning | Builds courses and assessments from source material and feeds them back as Q&A (C5) |
| SpotiTata music agent | Recommendation agent inside the demonstration mobile app (C5) |
3Connecting the assistants we already own
The review lists outbound connector recipes for these platforms, all served by one MCP server. The knowledge, tools and governance are identical; only the registration differs per vendor:
Microsoft Copilot Studio · Salesforce Agentforce · SAP Joule · ServiceNow · IBM watsonx Orchestrate · Google Vertex AI Agent Builder · Amazon Bedrock AgentCore · OpenAI AgentKit · Anthropic Claude · Oracle Fusion · any REST-capable framework.
Why this matters to us: we do not have to replace [existing_assistants]. A model's usefulness is limited by what it can reach, and this gives it governed access to our knowledge.
Governed by design
- Scope tokens: every call carries a token. Permitted tables resolve server-side, and a token with no scope is refused outright.
- One gate validates scope before any read, times the call and logs it.
- Per-tool cost weight and usage counters, so spending is visible per capability.
- Explicit no-results signal instead of a fabricated answer.
- Trace in every response: which tools ran, in order, with their inputs.
4Scores
| Category | Score |
|---|---|
| Agent portfolio breadth | 9.0 |
| MCP interoperability | 9.0 |
| Grounding | 9.0 |
| API / endpoint access | 8.8 |
| Architecture & extensibility | 8.8 |
| Security & governance | 8.7 |
| Tenant isolation | 8.7 |
| Observability | 8.5 |
5Objections the review answers
| Objection | Answer in the review |
|---|---|
| "Agents sound like a science project." | The business case is letting one question reach several systems (knowledge base, ERP, HR, a public API) without hand-writing every integration in advance. |
| "We are already buying Copilot." | Keep it. This is the knowledge layer underneath the assistants we already own. |
| "What stops an agent doing something wrong?" | Server-side scope validation, logging, and per-tool registration. Agent authority grows in stages: explain, recommend, prepare for approval, then act. |
| "MCP is new. Is it a risky bet?" | Small and reversible. MCP is one separate process over the same retrieval the REST API exposes. Stop it and everything else keeps running. |
6What our team should probe
| Probe | Our finding |
|---|---|
| Connect our target assistant to the MCP server end to end, and confirm the connector status for that specific vendor | [p_c4_1] |
| Call the REST API from one of our systems with a scoped token | [p_c4_2] |
| Define which actions an agent may take and which need human approval, before enabling any write actions | [p_c4_3] |
| Review the tool trace on two different questions and confirm different tools are selected | [p_c4_4] |
7Verdict, as stated by the review
Approved for production, grade A-. A portfolio of grounded, function-executing agents, connected to the wider agent ecosystem through one MCP server and reachable by any legacy or third-party system through a plain REST API.
Our assessment: [as_c4]. [as_c4_why]