Agents
An agent is an LLM-backed assistant you build, publish, and embed against your own governed data. Honeyframe agents come in two shapes:
- Simple agents — a system prompt, a model, an optional set of tools, and a knowledge base. The model decides when to call a tool and composes the answer.
- Structured visual agents — a visual flow of typed blocks (routing, LLM, tool call, Python, output, …) that you lay out explicitly, for deterministic, inspectable control over how the agent responds.
Both run through the same governed runtime: every SQL stays server-side, PII masking applies on every path, per-agent guardrails screen inputs and outputs, and every call is traced and costed.
Where agents live
| Surface | What it does |
|---|---|
| Agent Builder | Author a simple or structured agent — prompt, model, tools, knowledge, guardrails. |
| Agent Hub | Browse and manage published agents across the project. |
| Agent Chat | Talk to a published agent directly. |
| Agent Tools | Define the tools an agent may call. |
| Knowledge Banks | Vector stores an agent retrieves from (RAG). |
| Agent Review | Score an agent against a test suite of questions with judge traits (honesty, grounding, refusal-correctness). |
Guardrails
Every agent can carry a guardrail pipeline that runs on the query (input) and the response (output). Built-in guardrails include PII detection/redaction, numeric grounding (verify every figure against the query result), period-completeness caveats, and table-completeness disclosure. Guardrails degrade open — a guardrail failure never blocks a correct answer.
Tracing & cost
Each agent turn produces a structured trace (which blocks fired, which tools ran, latencies) and a token/cost ledger. Spend is visible per model in Administration → Monitoring.
See also
- Agent Tools reference
- Machine learning
- Catalog — the datasets and columns agents reason over