An AI agent is a software process that combines a language model with the ability to call external tools, take actions, and reason toward a goal across multiple steps. Unlike a single-turn chatbot, an agent can make multiple tool calls in sequence, use the results to refine its approach, and produce a structured output at the end.
What makes something an agent
- It calls tools or APIs to retrieve or act on real-world data
- It reasons about the results before deciding what to do next
- It can take multiple steps to complete a task
- It operates within a defined scope set by the user or operator
Agent hosts vs agent runtimes
An agent host is the application that surfaces the agent to a user (for example, Claude Desktop, ChatGPT, Cursor). An agent runtime is the underlying execution environment (for example, a custom Python script using an SDK, a serverless function, or a workflow orchestrator). Hoist Assets connects to both through the same Connector surfaces: MCP for hosts, REST API or CLI for runtimes.
Why agents make mistakes on Australian register questions
Language models have training data cutoffs and do not have real-time access to live registers. An agent that answers "is this asset encumbered?" from training data alone is guessing. The PPSR changes daily. ABN status, GST registration, and insolvency records change in near-real time. Agents need a live source connection to give reliable answers.
Why agents need this
The term "AI agent" matters for how Hoist Assets describes its audience. Hoist is not a dashboard product for human analysts (though humans can review Evidence Packs). It is designed for automated workflows where an agent initiates source checks, receives Evidence Packs, and routes the results: escalating risk flags to a human reviewer, continuing a workflow when checks pass, or surfacing human review items when the source data is ambiguous.
The distinction between "chatbot" and "agent" is relevant here because agents can act: they can trigger paid searches, create records, and pass results downstream. The Hoist Connector is designed with that in mind, including explicit cost-confirmation requirements before any paid search runs.
How Hoist uses this
Hoist Assets is designed for agent workflows. The MCP server exposes intent-level tools that agents can call without needing to understand PPSR or ABR APIs directly. Evidence Packs include structured fields (risk flags, human review items, confidence) that agents can parse and act on without unstructured text parsing. Hoist is model-agnostic: it works with any agent runtime that supports MCP, HTTP, or CLI subprocess calls.
What Hoist does not infer
Hoist does not know what model or provider is driving the agent. It does not track agent reasoning or store conversation context. It does not validate that the agent calling a tool is authorised to act on the result; authorisation is the responsibility of the operator configuring the agent. Hoist returns source evidence and flags; the agent and its operator decide what to do with it.
Related terms
- Hoist Connector: the MCP, API, and CLI surfaces AI agents use to call Hoist.
- Evidence Pack: the structured result Hoist returns to an agent after running source checks.
- Human review item: an action the Evidence Pack flags as needing human judgment, not just agent routing.
- MCP: the protocol most AI agent hosts use to connect to external tools like Hoist.
- Risk flag: a discrepancy in source data that an agent surfaces for attention.
- PPSR: Australian register that agents query through Hoist for security interest data.