Agent Team Scheduled Workflow with Gated Tool

Status: Preview

CRON scheduling and tool invocation gating both require a full Supabase instance. See Known Preview Surfaces.

This tutorial covers a more advanced automation pattern: an agent team runs a scheduled workflow that includes a write-class tool. Both the scheduled run and the tool call require separate owner approvals. Nothing happens without two explicit human decisions.

Prerequisites:

• Daily Workflow with Approval — understand single-agent scheduling first
• A workflow that includes at least one tool node with egress_class: write
• Two or more agents configured in your workspace

---

Step 1 — Create an agent team

lf team create \
  --name "Content Team" \
  --autonomy autonomous_with_gates

autonomous_with_gates allows the team to initiate runs but keeps all write-class gates active. The team cannot self-approve tool calls.

Add agents to the team:

lf team add-member <team-id> --agent <agent-a-id> --role executor
lf team add-member <team-id> --agent <agent-b-id> --role reviewer

---

Step 2 — Assign the workflow to the team

lf workflow assign <workflow-id> --team <team-id>

This creates an agents.workflow_assignments row linking the workflow to the team.

---

Step 3 — Schedule the workflow with approval required

lf schedule create \
  --workflow <workflow-id> \
  --assignee-type team \
  --assignee <team-id> \
  --cron "0 9 * * 1-5" \
  --timezone UTC \
  --require-approval

This schedules the workflow on weekdays at 09:00 UTC, assigned to the team, with a run-level approval gate.

---

Step 4 — Approve the scheduled run

At 09:00 UTC Monday–Friday, a pending team run appears:

lf approval list --type team_run

Approve it:

lf approval decide <run-approval-id> approve

The workflow engine begins executing. When it reaches the write-class tool node, execution pauses again.

---

Step 5 — Approve the tool invocation

The write-class tool call creates a agents.tool_invocations row with approval_status='pending'. The run is blocked until you decide.

List pending tool approvals:

lf tool-invocations list --status pending

Inspect what the tool would do:

lf tool-invocations inspect <invocation-id>

This shows the tool name, egress class, input payload, and the agent that triggered it.

Approve the tool call:

lf tool-invocations approve <invocation-id>

The tool executes. The workflow node completes and the run continues to the next node.

Reject the tool call:

lf tool-invocations reject <invocation-id> --reason "Unexpected payload"

The tool node fails with error_code='tool_invocation_rejected'. The workflow's failure policy (isolate by default) prevents the failure from cascading to other nodes. The run completes with a partial result.

---

Step 6 — Verify the audit trail

Each decision is recorded permanently:

-- Run-level approval decision
SELECT metadata->>'decision_by_lenser_id',
       metadata->>'decision_at',
       metadata->>'decision_reason'
FROM agents.agent_run_events
WHERE run_id = '<run-id>'
  AND event_type = 'approval_decided';

-- Tool invocation approval
SELECT approved_by, approved_at, rejection_reason, status
FROM platform.tool_invocation_logs
WHERE invocation_id = '<invocation-id>';

See Audit Trail Examples for full query patterns.

---

Why two gates?

Gate	What it blocks	Who resolves it
Run-level approval	The entire workflow run	Owner (human)
Tool invocation approval	A specific write-class tool call within the run	Owner (human)

The two gates are independent. You can approve the run but reject a specific tool call if the payload looks wrong. The workflow continues to other nodes after the rejection.

CRON cannot bypass either gate. A schedule set to requiresApproval: true always produces a pending run, even if the team has full_autonomy. The tool invocation gate is enforced separately by fn_invoke_tool() regardless of schedule settings.

---

Next steps

• Audit Trail Examples — full SQL patterns for both gate types
• Agent Boundaries — what agents can and cannot do at each autonomy level
• Approvals Reference — decision walkthrough with modify-and-approve