Fleets

Production flow

Use a fleet when the mission needs independently testable stages, deterministic routing, or approval. This support-resolution fleet triages risk, researches evidence, drafts a grounded answer, audits it, and sends unsafe or low-confidence work to a person.

Start from a template

When you create a fleet, the New fleet dialog offers prebuilt starting points in addition to a blank canvas. Selecting a template seeds the canvas with its nodes, edges, settings, and evaluation cases so you can begin from a known-good skeleton rather than an empty graph. You can also pull reusable node templates (agent nodes, control nodes) from the canvas palette to drop a pre-configured node onto the workflow.

1. 1

   Open New fleet

   Choose a template from the Starting point selector.
2. 2

   Review the seeded graph

   Confirm the nodes, edges, settings, and evaluation cases match your mission.
3. 3

   Customize and test

   Rename resources, adjust prompts, and run the fleet in Playground before publishing.

1234

1. 1Mission boundaryStart accepts the request; End exposes the final workflow result.
2. 2Risk routeStructured triage sends high-risk work directly to an escalation package and approval.
3. 3Resolution pathResearch, drafting, and audit are isolated agents that can be tested on their own.
4. 4Quality gateA deterministic condition sends approved work to End and exceptions to a human.

Canvas and nodes

Add nodes from the palette, drag their bodies to organize the canvas, and connect output handles to input handles. Use Zoom In, Zoom Out, and Fit View instead of browser zoom.

Node reference

Every node receives its predecessor's output as input and produces output for downstream nodes. The runtime executes in topological order, following edges based on outcomes.

Fallback branches

Every production node emits an outcome. When a node fails, the runtime follows the edge whose sourceHandle matches the outcome.

• Connect an unavailable branch from every agent and MCP node to a fallback agent or human review.
• Connect a verify_failed branch from every verify node to a correction agent or escalation.
• Connect a no_evidence branch from knowledge-retrieve nodes to web search or clarification.
• Enable require_fallbacks in fleet settings to block publishing when fallback branches are missing.

Configure a node

1234

1. 1Node paletteAdd agents and deterministic controls from one list.
2. 2Selected nodeThe canvas keeps the node and its neighbors visible while editing.
3. 3Execution contractSet the agent, prompt, response format, retry policy, and token budget.
4. 4SaveSave changes before testing; the header reports unsaved state.

Agent nodes can inherit fleet state or use a focused prompt. Choose structured output when the next condition needs fields such as requires_human or approved. Set retries only for transient failures and cap each node's token use. Configure a fallback model on each agent node so the fleet survives provider outages.

Test and audit

1234

1. 1PlaygroundRun a representative mission from the same fleet workspace.
2. 2ActivityWatch the active execution and approval state.
3. 3HistoryReview completed and failed executions.
4. 4Node traceInspect the ordered steps, branch decisions, timing, tokens, and node outputs.

• Run the normal path, every branch, one provider failure, and every approval path.
• Check Preflight before publishing -- it warns about missing fallbacks, responses, and knowledge.
• Use Activity for in-progress work and History for completed audit records.
• Keep the run ID when reporting an issue or correlating an external request.

Evaluate before publishing

Fleets support regression checks before publishing. Define a set of evaluation cases — inputs with an expected status and optional output checks — and Sentinel runs them through the fleet to catch behavior changes before they reach production.

1. 1

   Add cases

   In the fleet's configure panel, add cases with a name, input, expected success or failure, optional output-contains text, expected JSON, and a minimum-evidence requirement.
2. 2

   Set trials

   Choose how many times each case runs so you can catch flaky behavior.
3. 3

   Run evaluations

   Sentinel executes every case against the current graph and reports whether the fleet passed or failed.

• Cover the normal path, each branch, one provider failure, and every approval path.
• Use expected_output_contains and expected_json to pin down shape and content, not just status.
• Re-run evaluations after every edit so a regression is caught before publish.

Publish a fleet

1. 1

   Validate the graph

   Connect every required path, remove unreachable nodes, and give loops a hard iteration limit.
2. 2

   Set controls

   Define the token budget, retry limits, timeouts, fallback models, and approval gates.
3. 3

   Run representative missions

   Inspect node outputs and the durable run trace, including failure and approval paths.
4. 4

   Run evaluations

   Re-run the evaluation cases and confirm the fleet passes before publishing.
5. 5

   Save a version and integrate

   Use the fleet's Integrate panel for readiness, versions, and the scoped invocation key.