All guides

User Guide

Start with a workspace

Dashboard → organization and workspace selectors

Sign in to your account and select your company and workspace. Your role determines which actions are available. Changing workspaces changes the records you can see.

Demo workspaces contain synthetic data and use the simulator only. They do not run real models, tools, or billable tasks.

Agent Studio and playground

Agent Studio → agent → Edit / Playground

Create an agent draft with instructions, input variables, output format, and an output-token budget. Use Playground to test the draft with real model inference or the clearly labeled simulator.

Real inference is blocked when tools or knowledge require an unavailable approved gateway or index. Playground has no real tool execution. Never paste credentials into instructions or test input.

Versions and publishing

Agent Studio → agent → Versions; Workflow Studio → workflow → History

Saving changes creates or updates a draft; it does not make it production-ready. Publishing requires the normal evaluation and governance gates. Published versions are immutable.

Rollback restores a published snapshot to a draft. Review and publish that draft again to activate a new version.

Workflow builder

Workflow Studio → workflow → editor

Arrange declarative nodes and connect their inputs and outputs. Configure each node in its settings. Graphs are validated on save and publication; node settings are not executable JavaScript.

Manual triggers, model steps, switches, approval gates, and end nodes are supported. Model and external effects remain subject to release evidence and permissions.

Manual runs and approvals

Workflow Studio → published workflow → Run; Approvals → request

Start a manual run from a published workflow with the required input. The run pins its workflow and agent versions. Already-started runs retain their completion path when a commercial limit or expiration blocks new work.

Review the exact evidence before approving or rejecting. You cannot approve a run you started; a schedule or webhook creator or last editor also cannot approve their triggered run. Decisions are immutable and permissions are checked again when processing resumes.

Run monitoring and exports

Runs → run → trace / Plan & evidence

Filter runs by status and inspect steps, events, input/output evidence, token use, and reported or estimated cost. Waiting for approval and failed or unknown external outcomes are shown explicitly.

Authorized users can export tenant-scoped run CSV records. Exports redact structured credential fields and protect against spreadsheet formulas. A CSV export is not a full backup.

Templates

Templates → workflow templates → choose a template

Deploy a template bundle into your workspace as editable drafts. Review its inputs and configuration, then evaluate and publish through the normal gates.

Template acceptance tests use the simulator. CRM and accounting examples are mock/sandbox examples, not live provider integrations. The existing four-step templates onboarding tour remains available.

Evaluations

Evaluations → suite → run evaluation

Create structured cases and critical checks for the current draft. Run an evaluation job, inspect persisted case results, and review evidence. Passing critical results must match the exact draft, suite, and current policy before publication.

When a separate reviewer is required, the author cannot substitute their own review. Evaluation runs are not counted as real monthly workflow runs.

Describe-it builder

Agent Studio → Describe it (/app/build)

Describe the workflow you want in plain language. The suggested steps are laid out and validated on the server and saved only as drafts with no tools attached.

Review the draft, add approved configuration, evaluate it, and publish normally. A generated draft does not bypass any gate or carry out actions.

Schedules and webhooks

Limited

Workflow Studio → workflow → Schedule / Webhook

Schedule and signed-webhook definitions, deduplication, and checked dispatch are implemented. Published delivery and long-running autonomous model execution are not fully verified. Do not rely on these for unattended production processing without operational verification.

Delay nodes can park runs without holding a processing lease. Suspended or expired companies cannot start new scheduled or webhook runs.

Connectors and named integrations

Limited

Tools & Connections; Settings → Integrations

Custom REST definitions, typed actions, connection checks, revocation, and encrypted credential references are implemented. Outbound HTTP permits exact public IP-literal HTTPS destinations only; domain-name requests remain disabled pending a DNS-pinned transport.

Microsoft 365, Google Workspace, SAP, Oracle, Salesforce, ServiceNow, Jira, Slack, Teams, PostgreSQL, and MCP integrations are Planned. Knowledge indexing, BYOK model transport, SSO, and AI Shield are not yet available.