Concepts¶
Work items¶
A work item is one unit of work that a person would otherwise handle: a support ticket, a warranty claim, a loan application, a delivery exception. WIRE doesn't store your work items. It reads them through an input view that you define over your own tables, and writes agent results to an output table in your own database.
Use cases¶
A use case groups the agents, categories, and data for one business domain, for
example claims or underwriting. Agents, categories, and execution history are all
scoped to a use case.
Agents¶
An agent is a configured AI task. Each agent has:
- An agent type that sets its purpose and defaults (see below).
- A prompt that tells the model what to do.
- A model, such as
claude-haiku-4-5, served by Snowflake Cortex. - An input view and an output table.
- A version, so you can change an agent without losing the previous definition. Only one version of an agent is active at a time.
Agent types¶
| Type | Use it when | Default model |
|---|---|---|
| Triage | You need to classify work items into predefined categories with a confidence level. Your go-to for routing and prioritization. | claude-haiku-4-5 |
| Analysis | You need a deeper look at individual work items: summaries, timelines, risk assessments, and recommended actions. | claude-sonnet-5 |
| Summary | You want to improve fields on a work item, such as generating a concise title from a long description. | claude-haiku-4-5 |
| Rollup | You want to read many related rows, such as comments and worknotes, and keep only the important points. | claude-haiku-4-5 |
| Retrospective | You want to review other agents' output and assess quality, accuracy, or completeness. | claude-haiku-4-5 |
| Custom | None of the preset types fit. Every option is available and nothing is preset. | claude-haiku-4-5 |
| Common prompt | You have shared background or vocabulary that several agents in a use case should reference. It doesn't process work items on its own. | — |
Each type except Custom and Common prompt comes with a starter prompt that you can adapt.
Categories¶
Triage agents assign each work item one category from a list you define, such as
PRIORITY CLAIM, STANDARD CLAIM, or INELIGIBLE. Each category can carry a description
and examples, which WIRE includes in the prompt so the model applies them consistently.
Agent pipelines¶
Agents chain together through your data. A typical pipeline:
- A rollup agent condenses each work item's comments into key points.
- A triage agent categorizes every work item, using those key points as extra context.
- An analysis agent runs only on the items that triage marks as worth a deeper look.
You control the chaining in your input views: an analysis input view can join to the triage output table and filter on its category. Run agents in order, on a schedule, or one work item at a time.
Rules before AI¶
An input view can supply a pre-computed result in its RULE_ENGINE_RESULTS column. When
that value is present, WIRE skips the model for that row and records the rule result
instead. Use it for cases that you can decide deterministically, so you don't pay for AI
calls you don't need.
Change detection¶
Every input row has an UPDATED_DTM timestamp. WIRE processes only rows that changed since
their last result, so running an agent again doesn't reprocess, or pay for, unchanged work
items.
Execution tracking¶
Every agent run is recorded in control.agent_execution with its status, row counts,
token usage, errors, and the configuration it ran with. A run that finishes with some
failed rows reports COMPLETED_WITH_ERRORS rather than COMPLETED, so partial failures
are never silent. See Troubleshooting.
Humans in the loop¶
People can correct an agent's result. Triage results can be overridden with a different category, confidence, and reason; the original agent result is never changed, so you keep a record of both. People can also leave feedback on individual work items, which a rollup agent can summarize and feed back into future runs as additional context.