Steps, variables, triggers, schema
Agent Definition
The saved agent defines prompts, schedules, trigger rules, approved sources, tools, model choices, outputs, review gates, validation state, and schema-compatible JSON.
Agents
Agents
Repeatable AI work with approved sources, review, and evidence
Design agents in Clear Ideas, Claude Code, Codex, or other compatible tools. Run them in Clear Ideas with governed Sites, approved connections, metadata extraction, triggers, review gates, generated files, and signed evidence.
Governed agents
Turn Recurring AI Work Into a Reviewable Process
Clear Ideas Agents are saved processes for recurring AI tasks. They define the steps, variables, source scope, model choices, tool access, generated outputs, schedules, triggers, metadata extraction behavior, and review points that belong to each run.
The important difference is governance. Agents work from approved Site content and scoped connections, not broad file access or improvised tool use. The run record keeps source usage, tool calls, generated files, runner events, token usage, and evidence with the exact run that produced the result.
That makes agents useful for document review, diligence summaries, board reporting, compliance checks, metadata extraction, stakeholder updates, and generated deliverables where a team needs repeatability without losing oversight.
Agent surfaces
Build, Design, and Connect Agents
Start with the authoring path that fits the task, then govern the sources and tools the agent can use.
Sandboxed execution flow
How a Governed Agent Run Works
Clear Ideas separates the durable agent definition from the individual run. Each run receives a scoped contract, executes inside the sandbox, and returns reviewable evidence.
Least privilege
Scoped Run Contract
Each run receives only the Sites, files, credentials, connector tools, model routes, and egress permissions it is allowed to use.
Controlled runner
Sandboxed Execution
The runner executes steps with scoped credentials, durable state, callbacks, and controlled access instead of broad outbound freedom.
Inspectable outcome
Review and Evidence
Human approvals, warnings, generated files, tool calls, events, source usage, tokens, and evidence remain attached to the run.
What belongs to the agent
Definition, Access, Automation, Outputs, and Evidence
A useful agent packages the operating rules around the prompt so recurring runs can be inspected, reviewed, and improved.
- Governed RuntimeRun agents through sandboxed execution with scoped credentials, egress controls, structured runner events, resumable state, and evidence tied back to Clear Ideas.
- Agent Builder and DesignerBuild agents visually, generate them from plain-language descriptions, reverse-engineer them from strong examples, import JSON, edit them with an assistant, or author through MCP-compatible tools using the live schema.
- Scoped Source ConnectionsConnect agents to Clear Ideas Sites and approved MCP sources with read/write modes, tool allowlists, aliases, required connections, OAuth or token auth, and evidence settings.
- Triggers and SchedulesRun agents manually, on recurring schedules, from webhooks, when files upload, or as sub-agents invoked by another process.
- Metadata ExtractionUse agents in Site or folder-scoped extraction rules to populate structured metadata from approved documents, then merge or replace attributes as the record changes.
- Review and Human GatesUse human-in-the-loop pauses, variables that require override, validation warnings, structured outputs, benchmark checks, restart, cancel, continue actions, and run attention states for sensitive runs.
- Generated DeliverablesProduce final text, images, spreadsheets, documents, presentations, JSON, intermediate step outputs, and governed generated files attached to the run that produced them.
- Run EvidenceInspect streaming runner events, source usage, tool calls, tokens, credits, generated files, execution manifests, policy metadata, run traces, and signed evidence exports for agent runs.
- MCP Authoring and Run ControlUse MCP tools to inspect the live agent schema, validate definitions, create, update, clone, run, schedule, control, and retrieve run events from compatible authoring tools.
Visual definition
Build Agents Your Team Can Review
Visual design matters because sensitive AI tasks need clear steps, approved sources, generated outputs, and review points teams can inspect.
Visual builder
Map the Agent Before It Runs
The editor gives reviewers a concrete view of the process: steps, branches, Site scope, output templates, and the structure that will later produce run evidence.
Recurring work
Run Agents When Work Happens
Agents can start manually, on a schedule, from an upload-triggered extraction rule, from a Site or folder metadata rule, from an incoming webhook, or as a sub-agent called by another process. Routine work can move forward automatically while sensitive moments still pause for human input, approval, or correction.
Each run can chain prompts, goal-driven loops, webhooks, code steps, structured outputs, generated file steps, and approved Site content. Loops can keep every iteration or return the final value when the goal is met. The output is not disconnected from the process: Clear Ideas keeps the sources, tool use, generated files, review points, events, and evidence together.
- Limit AccessChoose the Sites, files, folders, metadata extraction scope, variables, and approved connections an agent can use before any work begins.
- Run the AgentExecute prompt steps, goal-driven loops, webhooks, code steps, sub-agents, structured outputs, and generated file steps on demand or from a trigger.
- Add ReviewPause for approval, check warnings, inspect source usage, and keep sensitive outputs from moving forward until the right person has reviewed them.
- Keep EvidencePreserve runner events, tool calls, generated files, token and credit usage, and signed evidence tied to the exact run.
Related reading
Keep Building Governed Agents
Evidence and operating guidance for governed agents.
- Agent Checklist: From Document Set to Repeatable ProcessA practical checklist for turning recurring document work into a governed process.
- Metadata-Powered Document SearchHow extracted attributes improve search, review, and downstream automation.
- Centralized Agent Operations Scale AutomationHow unified job tracking and extraction rules help automation scale.
- Agent ConnectionsHow approved sources, tool allowlists, read/write modes, egress controls, and evidence settings shape agent access.
- Agent DesignerHow prompts, Codex, Claude Code, and MCP clients help shape governed agents in Clear Ideas.
- Clear Ideas + Claude Code via MCPHow external AI clients can work with governed Clear Ideas context through MCP.
- Clear Ideas + Codex via MCPHow Codex can use Clear Ideas context while preserving governed access.
Frequently Asked Questions
What is a Clear Ideas Agent?
A Clear Ideas Agent is a saved process for recurring AI work. You can define the steps, inputs, approved Sites, connections, output formats, review points, and triggers once, then run the agent whenever the same work comes up again. Common examples include document review, diligence summaries, board reporting, compliance checks, extraction, and file generation.
What makes Clear Ideas Agents different from generic AI agents?
Clear Ideas Agents are built for reviewable execution: which files were allowed, which tools were available, which steps ran, where a human needed to review, what files were generated, and what evidence remains for audit or client review. That makes Agents better suited for sensitive business processes than one-off prompting.
Can agents use external systems?
Yes. Agents can use approved connections, including Clear Ideas Sites and supported MCP providers. Administrators can control read/write mode, allowed tools, aliases, required connections, authentication, source instructions, egress, and evidence settings so external access is deliberate instead of improvised.
Are agent runs sandboxed?
Yes. Agent work runs through sandboxed execution with a scoped run contract. Agents do not receive general outbound access by default; destinations are granted through approved connectors, model endpoints, Clear Ideas services, and explicitly enabled web tools.
Can agents run without manual intervention?
Yes. Agents can be started manually, scheduled, triggered by webhooks, triggered when files are uploaded, or called by another agent. You can still add human approval steps for high-risk moments, so automation can move routine work forward while keeping sensitive decisions under review.
Can agents extract metadata from Site content?
Yes. Site metadata extraction rules can run approved agents against a whole Site or a specific folder. The rule controls whether extracted attributes replace or merge with existing metadata, and the run record keeps status and evidence attached to the work.
What can an agent produce?
Agents can produce final responses, structured JSON, extracted values, intermediate step outputs, images, spreadsheets, documents, presentations, and generated files that stay connected to the run that created them. That is useful when the deliverable needs to be reviewed, reused, downloaded, or included in an evidence export.
How do reviewers check what happened?
Reviewers can inspect the run record, including steps, source usage, tool calls, generated files, runner events, token and credit usage, attention states, and evidence exports. Teams can review the path the agent took before they reuse, share, or approve the output.
Ready to get started?
Share sensitive information securely with clients, auditors, and partners. Then turn approved content into cited answers, Governed Agents, and measurable engagement.
Start Free
Book a DemoNo credit card required
Need help?
Get personalized assistance
Speak with our sales team to find the perfect plan for your organization.
Technical support & resources
Access support resources, documentation, and help guides.