Skip to Content

Agent Studio

Agent Studio is the visual orchestration canvas in AIX Console for designing, configuring, testing, and deploying intelligent agents.

Core Components of an Agent

Every agent configured in Agent Studio is composed of five declarative layers:

┌────────────────────────────────────────────────────────┐
│                   Agent Configuration                  │
├────────────────────────────────────────────────────────┤
│ 1. Identity & Role   (Name, Description, Tone)         │
│ 2. Foundation Model  (Qwen3, DeepSeek, Claude, Llama)   │
│ 3. System Prompt     (Boundaries, Few-shot examples)   │
│ 4. Tools & MCP       (Connected databases, APIs)       │
│ 5. Memory & Context  (Session memory, Vector retrieval) │
└────────────────────────────────────────────────────────┘

1. Identity & Role

  • Display Name: Shown in the agent directory, playground, and audit logs.
  • Mode:
    • Reactive: Answers direct user inquiries without unprompted background tasks.
    • Collaborative: Coordinates with peer agents and human approvers on complex workflows.
    • Proactive: Runs scheduled automations, listens to webhook events, and emits notifications.
  • Enterprise Autonomy & Policy Binding: The deployment mode dictates when and how the agent is triggered. Authority boundaries and risk gates (L1 Full Auto, L2 Human-in-the-Loop, L3 Human-in-Command, L4 Circuit Breaker) are governed by AEP policies. See AEP Governance & 2D Autonomy Matrix.

2. Model Selection & Hyperparameters

  • Model Router: Choose from models hosted in your tenant’s cluster or external enterprise endpoints.
  • Temperature: Range 0.0 to 1.0. Keep low (0.1 - 0.3) for deterministic code generation and compliance; raise (0.7 - 0.9) for creative copywriting.
  • Max Tokens: Guard against runaways by setting an explicit per-turn response limit.

3. System Prompt Engineering

The system prompt is injected at the root of every inference call. Follow the ROLE-TASK-CONSTRAINT format:

# Role
You are a senior data analysis copilot for enterprise merchandising teams.
 
# Tasks
1. Parse user queries for inventory and sales trends.
2. Formulate analytical SQL queries against the read-replica database.
3. Summarize findings into executive bullet points.
 
# Constraints
- NEVER perform write or delete operations on tables.
- If data confidence is below 80%, explicitly state the limitation.
- Format all currency in standard local format (e.g., ¥ or $).

4. Mounting Tools via MCP

In the Tools tab, browse through tools provided by your connected Model Context Protocol (MCP) servers:

  • Check the tools your agent is permitted to call.
  • Specify tool call timeouts and fallback strategies.
  • Inspect JSON schemas for all input parameters.

5. Memory & Context Injection

  • Session Window: Retains recent user messages within the current conversational thread.
  • Knowledge Store (RAG): Connect vector indexes from AIX Knowledge Hub for semantic document retrieval before generating responses.

Deployment & Versioning

When saving an agent, Agent Studio automatically captures an immutable configuration snapshot:

  • Compare diffs between revisions in /dashboard/agents/[id]/versions.
  • Instantly roll back to previous stable versions with 1-click.