Skip to main content
Version: 0.9.19

Key concepts

This page introduces the core building blocks of AlphaAgent. Each concept has a fuller chapter in the Studio guide or platform internals; this is the orientation you need before diving in.

Agents​

An agent is an AI assistant you define: its instructions (system prompt), its model settings, and the data sources and knowledge it can draw on. Agents are versioned — every time you edit an agent, AlphaAgent saves a new version rather than overwriting the old one, so you always have a history you can return to.

A separate active configuration decides which version of an agent is currently live, along with which knowledge graph, connectors, and execution environment that live agent uses. Editing an agent never changes what is running until you activate it. Learn more in Agents.

Swarms​

A swarm is a team of agents that work together: a supervisor agent receives the request, plans, and delegates parts of the work to specialist agents. Swarms are useful when a task spans several domains and benefits from dividing the work. See Swarms.

Knowledge Graphs (AMPG)​

A Knowledge Graph, also called an AMPG (Augmented Multi-Resolution Property Graph), is built from documents you upload. AlphaAgent processes those documents into a graph of connected information that agents can retrieve from when answering questions. The graph is stored in your own knowledge-graph database. See Knowledge Graphs (AMPG).

Data connectors​

A data connector links an agent to an external data source. Supported types include SQL databases (PostgreSQL, MySQL), REST APIs, external tools (MCP servers), Snowflake, and AWS data sources. Connectors are versioned, and an agent's active configuration pins the exact connector versions it uses. See Data connectors.

Execution environments​

An execution environment is the sandboxed compute where an agent runs code. You can use a prebuilt environment or define a custom one with specific libraries and resource settings (memory, storage, and time limits). Each agent's active configuration references the environment it runs in. See Execution environments.

Code interpreter and datasets​

Agents can run Python and shell code in their execution environment to analyze data. Each conversation gets its own isolated workspace for files. When an agent produces a result table, it registers it as a dataset — a reference that lets the chat show a compact preview, chart, or grid instead of dumping every row into the conversation. See Code interpreter and datasets and, for the technical detail, Datasets and data flow.

Chat modes​

When you chat with an agent, you can use one of two modes:

  • Express — the fast, direct path for straightforward questions and quick tasks.
  • Deep Traversal — a planner-first path that draws on knowledge-graph retrieval and can present a plan for your approval before doing deeper, multi-step work.

See Chat.

Workflows​

A workflow is a multi-step pipeline built as a graph of nodes — task nodes (run an agent), document-analysis nodes, conditional nodes, and handoff nodes (pause for human input). Workflows can run on demand or on a schedule that you describe in plain language. See Workflows.

Licenses​

A license key is the entitlement, created in the Console, that authorizes one Studio deployment. When you deploy Studio you provide a one-time license token; the deployment then activates the license and sends periodic health and usage signals back to the Console. See License keys and Licensing and telemetry.

How they fit together​

These pieces stack up like this: you create agents (optionally grouped into swarms), give them connectors and knowledge graphs for context, and let them run code in execution environments. You interact through chat or automate them with workflows, and the whole thing is authorized by a license. The next page, How it works, walks the full journey end to end.