Documentation

Agents Overview

View as Markdown

Agents turn recurring document-grounded work into controlled, repeatable AI processes. Instead of re-running the same prompt manually, you define the Agent once, connect it to approved sources, run it on demand or on a schedule, and keep a governed record of each run.

Agents are the execution layer in the Clear Ideas model:

  • Sites hold approved content, collaborators, permissions, analytics, and evidence.
  • AI answers and searches from approved context.
  • Agents run repeatable tasks over that approved context and produce reviewable outputs.

What Agents Are Good For

Agents are especially useful when you need to:

  • summarize, review, or transform approved documents the same way every time
  • chain multiple AI steps into one controlled operating procedure
  • run work from schedules, upload triggers, webhooks, or another Agent
  • generate spreadsheets, documents, presentations, images, or structured outputs
  • enforce source limits, tool policy, review gates, and evidence requirements
  • benchmark and improve output quality over time

Core Agent Building Blocks

An Agent can include:

  • prompt steps
  • loop or container steps
  • structured output schemas
  • output templates
  • Agent variables and trigger variables
  • sub-agent execution
  • model and reasoning settings
  • generated file outputs
  • source connections
  • controlled tools and egress settings
  • optional web access, when allowed by policy

Model and Reasoning Settings

Agents use the same model selector pattern as AI Chat, with one important difference: Agents often run later. For scheduled or future runs, model selectors may include latest aliases in addition to concrete model versions.

Use a latest alias when you want future runs to follow Clear Ideas' current recommended version. Choose a concrete model version when repeatable behavior across runs, benchmarks, or governed comparisons matters more.

Agent-level model and reasoning settings act as defaults. Individual prompt steps can override them when one step needs a faster model, a more capable model, a private model, or a different reasoning effort.

Runs, Evidence, and Generated Files

Each Agent run records what ran, when it ran, the source context used, the step outputs, generated files, and the final result. Organization administrators can review governed records and export evidence from Audit & Governance > Governance Activity when policy requires audit or review.

Generated spreadsheets, documents, presentations, and images stay connected to the run that produced them.