Generative AI
A Zeta Alpha agent is a language model working in a loop with tools. The model itself runs on an inference endpoint you choose: OpenAI, Azure, Anthropic, Bedrock, or your own server such as vLLM. Everything around the model is the agent harness, the program in the Agents SDK that prepares each model call, runs the tools the model asks for, and returns the answer.
How a Turn Runs
- A message arrives from the chat UI, the Chat API, or an integration such as Slack or Microsoft 365, with optional context: the documents the user is looking at, active filters, the conversation so far.
- The harness builds the prompt. It combines the agent's configured instructions, the catalog of available skills, recalled memories and the context into one system prompt.
- Tool rules decide which tools this call offers, and which of them must ask the user before running.
- The harness calls the model with the prompt, the conversation and the tool list. The model either answers or asks for one or more tools to run.
- The harness runs the requested tools, for example a search in your index, reading a document, or a call to an MCP server, and passes the results back to the model. Steps 3 to 5 repeat until the model answers without calling a tool, up to a configured limit.
- The harness finishes the answer. It resolves citations to the documents they point at and streams the answer to the user while it is written.
Skills follow the same path. A skill is a markdown document of instructions: the prompt lists each skill's name and description, and when the model decides a skill applies, it loads the full text through a tool.
The model never reaches your data, tools or users directly. It sees only what the harness puts in the prompt and the tool results, and every action goes through the harness.
Why a Harness
A model endpoint takes a prompt and returns text. A product needs more: grounded answers with citations, conversations that survive a restart, history that fits the model's context window, and control over what the agent may do. The harness does that engineering once, for every agent, and different configuration turns it into a research assistant, a document Q&A bot or a domain workflow without new code.
| Capability | What you get | Docs |
|---|---|---|
| Composable capabilities | Everything below is a pluggable source, enabled per agent with include/exclude filters: behavior is configured, not coded | Concepts |
| Retrieval and citations | Search, browse, and read any Zeta Alpha index, with resolved citations linking back to sources | Tools |
| Knowledge capture | Notes, tags, and private document ingestion through conversation | Note Tools |
| Data analysis | CSV/Excel and SQL exploration, aggregation, and inline charts | DataFrame Tools |
| Web access | Fetch and read public pages and PDFs, hardened against hostile content | Web Tools |
| Skills | Markdown-defined behaviors the agent discovers and applies on demand | Agent Skills |
| Memory | Facts persisted across sessions and loaded into the prompt | Memory |
| Multi-agent | Anonymous sub-agents and named specialist delegation | Delegation |
| MCP | External tool servers, including per-user OAuth authorization | MCP Tools |
| Model providers | OpenAI, Anthropic, Azure, Bedrock, Ollama, and any OpenAI-compatible endpoint (vLLM), with tool-calling accommodations for small models | LLM Configuration |
| Context management | Keeps as much history as fits by virtualizing old and large tool responses before older answers, so smaller and larger models swap in without code changes | Context Window Management |
| Resilient conversations | Server-owned turns that survive restarts: reattach from any replica, resume a crashed turn from its checkpoint | Conversations · Persistence |
| Streaming | SSE streaming with reconnection, turn discovery, and cancellation | Streaming |
| Tool streaming | Templated progress text while tools run, and per-tool shaping that keeps large results out of model context while the UI gets the full result | Tool Events |
| Testing and evaluation | Behavior specs, and RAGElo judging through the same provider layer, so the judge can be any OpenAI-compatible backend | Testing |
| Deployment | Run on the Zeta Alpha platform or self-host the REST API | Self-Hosted |
Extending the Harness
Every step of the turn can be changed, by configuration first and by code where configuration is not enough. Each capability in the table above is a pluggable source: you can turn it off, or add your own beside it. Beyond capabilities:
| Extension point | What it changes | Docs |
|---|---|---|
| Tool rules | Hide tools, require the user's approval, fill in arguments, or replace or withhold a tool's result | Tool Rules |
| Dispatch rules | Answer some requests directly, without calling the model | Dispatch |
| Interjections | Add supervisor messages between model calls, such as a reviewer or a warning after repeated tool failures | Interjections |
| Message processing | Transform the answer before it reaches the user | Message Processing |
| Storage | Keep conversations and turn state in your own database or object store | Persistence |
| Custom agents | Write your own agent in Python and reuse any of the above | Custom Agent Development |
How these pieces fit together internally is in Agents SDK Architecture.
There are two ways to work with agents, depending on what you need: configure them, or call them from your application.
Configure Agents
Define agent behavior through JSON configuration: search tools, instructions, skills, memory, and sub-agents. No application code required for most use cases.
Start here if you want to:
- Create or customize agents for your domain
- Configure what tools, memory, and instructions an agent uses
- Deploy agents to the Zeta Alpha platform
- Test and debug agent behavior locally
Chat API
Call deployed agents from your application through REST or streaming endpoints. Send messages, pass context, and handle responses.
Start here if you want to:
- Integrate agent conversations into your product
- Use the streaming or REST Chat API
- Pass custom context or handle function call requests