Running & Managing

This chapter shows how to keep agents humming in production: monitor health, automate schedules, process bulk tasks, manage spend, and get alerted when something drifts.

8.1 Live Task Dashboard

Panel	What It Shows	Why It Matters
Task List	Every job with status badge (Queued ▢, Running ▶, Succeeded ✔, Failed ✖)	Spot stuck or failing runs instantly
Timeline	Gantt‑style trace of each tool call with latency bars	Identify slow MCP hops (> 200 ms)
Agent Chat	Real‑time stream of model thoughts (when debug mode on)	Understand reasoning; copy snippets for support tickets
Details Pane	Inputs, outputs, cost, sandbox ID, policy decisions	Forensic debugging & auditing

Access via Home → Tasks or Builder → Tasks → View Runs.

Anomaly Highlight 🔍 – Rows turn amber if runtime > P95, red if policy blocked.

8.2 Recurring Schedules

Why schedule? Eliminate manual triggers, ensure reports arrive before business hours, and smooth out workload spikes.

Schedule Type	Cron DSL	Typical Use
Hourly	0 * * * *	Sync CRM leads
Daily	30 6 * * *	6 : 30 AM KPI digest
Weekly	0 1 * * 1	Monday 1 AM log rotation
Monthly	0 2 1 * *	First‑of‑month invoice summary
Custom	Any Cron‑5	Edge cases (e.g., 15th & last day)

How‑To: Builder → Tasks tab → flip Run on Schedule → pick template or enter Cron → save. Scheduled tasks appear with a ⏰ icon and green Next Run column.

Pro Tip ✨ Start with a small test window (e.g., next minute) before committing to production cadence.

8.3 Bulk Task Queues

Need to process 10k PDFs or 50k leads? Use Bulk Upload.

1. Prepare CSV with a header row matching input names (file_url, customer_id, …).

2. Builder → Tasks → Bulk Run → upload CSV.

3. Choose concurrency (default 50) and dead‑letter policy (retry×3 or move to Failed queue).

4. Click Launch – progress bar shows completed vs remaining.

Scale Math: Each micro‑VM sandbox can handle ~2-3 tool calls/sec. A 10k job with avg 3 calls = ~25 min at 50 concurrency.

8.4 Cost & Usage Dashboards

Navigate: Home → Cost.

Chart	Metric	Actionable Insight
Tokens Burned	per agent / per model	Tune prompts, choose cheaper model
MCP Latency P95	by tool	Identify slow back‑ends (DB, SaaS)
Sandbox Seconds	compute per integration	Spot heavy workloads (e.g., large GPT‑4 calls)
Spend vs Budget	daily burn vs set budget	Auto‑throttle when 80 % reached

Budget Guardrails 💰 – Set Hard Cap (terminate runs) or Soft Cap (pause schedules, alert owner). Budget events are logged in Policy journal.

8.5 Alerts & Notifications

Trigger	Channel	Configuration
Task Failure (> N retries)	Slack DM, Email	Toggle in Org Settings → Alerts
Budget Hit 80 %	Slack #ops‑alerts	Enabled by default
Policy Block	Webhook POST	Custom endpoint
Latency Spike (> 500 ms)	PagerDuty	Add via Integrations

Alerts include deep links to Task ID and Timeline for quick triage.

8.6 Operational Best Practices

1. Tag Agents by business unit (finance/, support/) – dashboards auto‑group.

2. Stagger Schedules (± 5 min) to avoid thundering‑herd on DB.

3. Enable Must‑Cite on KB tools to catch hallucinations early.

4. Review Policy Journal weekly; look for repeated blocks (might indicate missing tool scopes).

5. Export Metrics via OpenTelemetry to Grafana or Datadog.