SSentinelDocumentation
Open Sentinel
Example: build one production-level fleet

Examples

Example: build one production-level fleet

Build, test, audit, and publish a support-resolution workflow with structured gates and human review.

Mission and architecture

Production Support Resolution converts a customer request into either a grounded, audited response or a documented human-review package. It uses four deployed specialists and two deterministic quality gates.

End-to-end production support fleet zoomed to fit
End-to-end flow: triage -> risk route -> research -> draft -> audit -> quality route, with human approval on both exception paths.
  1. 1Triage contractThe production agent from the previous example emits requires_human.
  2. 2High-risk branchCreate an escalation package and pause for approval.
  3. 3Grounded responseResearch and drafting stay separate so evidence can be inspected.
  4. 4Audit gateOnly approved output reaches the end automatically.

Build the canvas

  1. 1

    Create the boundary

    Create the fleet, then place Mission Start and Mission End.
  2. 2

    Add specialists

    Add Support Intake & Risk Triage, Evidence Research Specialist, Grounded Resolution Composer, and Policy & Safety Auditor.
  3. 3

    Add controls

    Add two If/Else nodes, an escalation-package transform, and a Human Approval node.
  4. 4

    Arrange and connect

    Drag nodes into reading order, connect every output, connect both condition branches, then use Fit View.

Configure contracts

Production fleet selected auditor node and configuration panel
Each selected node exposes its own prompt, response format, retry rule, and token ceiling.
  1. 1Agent assignmentChoose the already-tested production specialist.
  2. 2Focused promptDescribe only this stage's responsibility and input.
  3. 3Response formatUse structured output when a later node reads fields.
  4. 4Reliability limitsBound retries and tokens so failure is visible and cost is controlled.
  • Triage emits requires_human and confidence.
  • The first condition branches on requires_human.
  • The auditor emits approved, issues, and required_changes.
  • The final condition branches on approved.
  • Both exception paths reach Human Approval; no branch is unreachable.

Test in Playground

Open Playground in the fleet header. Run one routine request, one security-sensitive request, one missing-evidence request, and one response that the auditor must reject. Use Activity while the run is live to see the current node and any approval pause.

Audit and publish

Successful production fleet history entry and node execution trace
A representative successful run completed eight recorded steps within the 2,400-token fleet budget.
  1. 1Run recordStatus, start time, duration, and total tokens establish the outcome.
  2. 2Execution orderEvery agent and condition appears in the order it ran.
  3. 3Step evidenceOpen a step for inputs, structured output, errors, timing, and token use.
  4. 4IntegrateAfter readiness checks, save a version and create a scoped invocation key.
  1. 1

    Inspect History

    Confirm the first wrong step--not only the final answer--and compare branch fields with the node schema.
  2. 2

    Save a version

    Record the tested configuration so later edits do not erase the known-good release.
  3. 3

    Check readiness

    Confirm deployed agents, complete paths, budgets, approvals, and representative runs.
  4. 4

    Integrate

    Use the generated endpoint and scoped key; retain the returned run ID for audit and support.