Platform-Managed Agents

Build agents using the platform's built-in LLM runtime

Platform-Managed Agents

This guide covers building agents that run entirely on the platform's built-in LLM runtime. The platform handles inference, tool calling, memory, and tracing end-to-end — you just configure the agent.

When to Use Platform-Managed Mode

Choose platform-managed when:

• You want to get started quickly without writing code
• Your agent's logic can be expressed through a system prompt + tool access
• You need built-in memory, guardrails, and observability out of the box
• Your use case is primarily conversational (Q&A, support, document lookup)

Choose external pro-code instead when:

• You need custom reasoning loops (graph-based, multi-step planning)
• Your agent requires complex state machines or branching logic
• You're building multi-agent orchestration with specific coordination patterns

Creating a Platform-Managed Agent

Step 1: Create Agent in Studio

Navigate to Agent Studio > Agents > New Agent and configure the Identity tab:

Step 2: Configure the System Prompt

The system prompt is the most important part of a platform-managed agent. It defines the agent's personality, capabilities, and constraints.

Best practices for system prompts:

You are [Role Name], an AI assistant for [Organization].

## Your Purpose
[1-2 sentences describing what you help with]

## Your Capabilities
- [Capability 1 with tool reference]
- [Capability 2 with tool reference]

## Rules
- Always [constraint]
- Never [restriction]
- When uncertain, [fallback behavior]

## Response Style
- [Tone: professional/casual/technical]
- [Format: bullet points, paragraphs, tables]
- [Length guidance]

Example:

You are the HR Policy Assistant for ACME Corp.

## Your Purpose
Help employees find answers to HR policy questions by searching the company knowledge base.

## Your Capabilities
- Search the HR knowledge base for policy documents
- Summarize complex policies in plain language
- Create support tickets for HR-specific requests

## Rules
- Always cite the specific policy document when answering
- Never provide legal advice — refer to Legal team for legal questions
- When you can't find an answer, create a ticket for the HR team
- Respect data privacy — never share employee personal information

## Response Style
- Professional but friendly tone
- Use bullet points for multi-part answers
- Keep responses under 300 words unless the user asks for detail

Step 3: Select a Model

In the Model tab:

Model selection guidance:

• High accuracy required (legal, financial): Claude Sonnet or GPT-4o
• High throughput / cost-sensitive: Llama 3.3 70B or Gemini Flash
• Reasoning-heavy: Claude Sonnet or o3-mini
• Simple classification/routing: Llama 3.1 8B or Gemini Flash

Step 4: Assign Tools

In the Tools tab, grant access to the MCP tools your agent needs:

• Knowledge Base Search — for RAG-based document lookup
• Create Ticket — for escalation workflows
• Web Search — for real-time information
• Code Execute — for data analysis (requires semi-autonomous or higher risk level)

Each tool can have conditions:

• Rate limits — max invocations per conversation or per hour
• HITL requirement — require user approval before the tool executes
• Allowed parameters — restrict tool input values

Step 5: Configure Memory

In the Memory tab:

For most agents, start with short-term only and add long-term memory once you understand the usage patterns.

Step 6: Set Access Control

In the Access tab:

• Visibility: public (all users), restricted (specific groups), private (dev only)
• Allowed Groups: Select IdP groups that can access this agent
• Allowed Roles: consumer, builder, admin

Step 7: Test in Playground

Before deploying, test your agent in the Playground (Studio > Playground):

1. Select your agent from the dropdown
2. Send test messages covering your key use cases
3. Verify tool usage, response quality, and guardrail compliance
4. Adjust the system prompt or model settings based on results

Step 8: Deploy

In the Deployments tab:

1. Click Create Deployment
2. Choose target environment (staging → production)
3. Optionally enable canary mode (routes a % of traffic to the new version)
4. Monitor metrics in the Monitoring dashboard

Testing with the Eval Suite

Create golden test sets to systematically validate your agent:

1. Navigate to Studio > Evals
2. Create a test suite with input/expected-output pairs
3. Run the eval — the platform scores responses using LLM grading
4. Track eval scores across versions to prevent regressions

See the Eval Suite & Playground Guide for details.

Monitoring

Once deployed, monitor your agent in Studio > Monitoring:

• Success rate — % of conversations completed without errors
• Latency P95 — response time at the 95th percentile
• Cost per day — LLM token costs
• HITL escalation rate — % of conversations requiring human approval
• Tool usage — which tools are called most frequently

Common Patterns

RAG-Based Q&A Agent

System Prompt: "Search the knowledge base before answering. Always cite sources."
Tools: Knowledge Base Search
Memory: Short-term only
Model: Claude Sonnet (high accuracy)

IT Support Triage Agent

System Prompt: "Classify issues, search KB, create tickets for unresolved items."
Tools: Knowledge Base Search, Create Ticket
Memory: Short-term + Long-term (remember past issues)
Model: Llama 3.3 70B (cost-effective for classification)

Research Assistant

System Prompt: "Search the web and knowledge base, synthesize findings."
Tools: Web Search, Knowledge Base Search, Text Summarizer
Memory: Short-term + Long-term
Model: Claude Sonnet (best for synthesis)

Next Steps

• Tool Integration Guide — register custom MCP tools
• Knowledge Bases Guide — create and populate knowledge bases
• Memory Configuration Guide — configure the 4 memory types
• Agent Design Patterns — when to use which execution mode