Skip to main content

How the Agent Harness Works

The SDK ships a production-ready agent harness. You don't build an agent from scratch - you configure one. The same agent code with different configuration becomes an internal research assistant, a document Q&A bot, or a domain-specific workflow executor.

This page gives you the mental model and shows you what is possible so you can find the right guide quickly.

The Core Idea​

The agent assembles itself at runtime from capabilities - plug-in components you toggle on and off through configuration. You pick which capabilities are active, configure how they behave, and the agent does the rest. No Python required for most use cases.

You manage capabilities through the CLI:

za agents capabilities tool-sets list           # see available tool sources
za agents capabilities tool-sets configure web-tools  # enable web browsing
za agents capabilities memory-stores settings    # configure memory
za agents show                                   # see the full picture

What the Agent Can Do​

Search and Retrieve​

Connect the agent to your Zeta Alpha knowledge base. It can search, browse, filter by facets, read full documents, and cite sources with links back to the original.

za agents capabilities tool-sets configure index-tools

→ Configuring Tools

You can also give each user access to their own private uploaded documents:

→ Configuring Tools - User Documents

Analyze Data​

The agent can load CSV and Excel files, explore schemas, filter, aggregate, join tables, and produce charts - all through natural language.

→ Configuring Tools - DataFrame Tools

For SQL databases, the agent explores schemas, queries data, and produces results without writing raw SQL:

→ Working with SQL Databases

Browse the Web​

Enable web-tools and the agent can fetch any public URL, extract clean content, and incorporate it into its responses.

za agents capabilities tool-sets configure web-tools

Learn New Behaviors via Skills​

Skills are markdown documents that teach the agent how to handle specific tasks. Write a skill file, drop it in the skills/ directory, and the agent discovers it automatically:

---
name: summarize-document
description: 'Summarize a document. Use when asked to "summarize" or "give me a summary".'
---

## Steps
1. Retrieve the document using the document context
2. Produce a structured summary: key findings, methodology, conclusions
3. Cite specific sections

Skills can also be authored as tag-notes in the platform UI - no deployment needed. And users can create new skills through conversation: the agent writes a proper SKILL.md from a natural language description.

→ Using Agent Skills

Remember Across Sessions​

The agent saves facts worth remembering - user preferences, project context, recurring corrections - and loads them into the system prompt on the next session.

→ Configuring Memory

Delegate to Specialist Agents​

The agent can spawn anonymous sub-agents for parallel work or route tasks to named specialist agents. Delegation is configured, not coded:

{
  "in_memory_delegable_agents_source_configuration": {
    "enabled": true,
    "agents": [
      { "agent_identifier": "compliance-checker", "description": "Reviews content for regulatory compliance." }
    ]
  }
}

→ Configuring Delegation · Sub-Agent (anonymous)

Connect External Tools via MCP​

Any tool with an MCP server - GitHub, Slack, databases, internal APIs - can be connected to the agent:

za agents capabilities mcp-servers add

→ Using MCP Tools

Understand User Context​

The agent can see what the user is currently looking at - the selected document, active search filters, the current tag - and use that as implicit context. No need for the user to repeat themselves.

→ Configuring Context

Control the System Prompt​

Assemble the system prompt from modular instruction sources: date/time awareness, citation formatting rules, onboarding guidance, platform self-knowledge, and free-text behavioral instructions.

→ Configuring Instructions

Going Deeper​