Skip to Content
Agentic WorkspaceTask Agent10. Monitor & iterate

Monitor & iterate

By the end of this page you will know how to monitor your Task Agent’s production performance, inspect individual runs, track costs, and use insights to improve your agent.

Prerequisites

  • A deployed Task Agent (Deploy).
  • Admin or role.

Observability overview

Monitoring lives at the deployment level. Open Deployments in the project sidebar and click a deployment to open its detail view. The left navigation groups the tools into sections:

SectionPages
OverviewDashboard, Versions
MonitoringObservability, Logs
RuntimeAgents, Executions, Schedules, Triggers, Conversations
AccessAPI Keys, Embed
ConfigureSettings

The Dashboard opens first, with Quick Metrics (24h): Requests, Error Rate, P95 Latency, and Unique Users, plus a What’s Deployed panel and the deployment’s access details (URL, authentication).

Observability dashboard

Open Monitoring → Observability for live metrics over a selectable time range (Last 1 hour by default, with Summary / Traces / Activity views):

MetricDescription
RequestsTotal requests in the window, with period-over-period change
Avg LatencyAverage end-to-end response time
Error RatePercentage of failed (5xx) requests
APDEXApplication performance score (1.00 = Excellent)

Below the metrics are a Request Rate Over Time chart, Top / Slowest Endpoints, and a Recent Errors feed. Resource (CPU/Memory) metrics appear when the workspace exports them via OTEL.

Viewing executions (runs)

Open the Executions list

In the deployment detail view, go to Runtime → Executions. This is the chronological list of agent runs, with a status filter and columns for:

  • Time — When the run occurred
  • Agent — Which agent handled the request (with its type badge)
  • Status — Completed, Failed, or in-progress
  • Duration — Total execution time
Deployment Executions page listing agent runs with time, agent, status, and duration

Inspect a single execution

Click a row to expand its details inline:

  • Run ID — The unique identifier for the execution (copyable).
  • Tokens — LLM tokens consumed by the run.
  • LLM Calls — Number of model invocations.
  • Tool Calls — Number of tool executions.
  • View Conversation — Opens the full conversation/trace for the run.
Expanded execution row showing Run ID, token count, LLM calls, tool calls, and a View Conversation link
💡

The Conversations page (Runtime → Conversations) tracks chat sessions for Copilot deployments — total sessions, active sessions, unique users, average duration, and per-session message counts.

Tracking costs

Each run displays its credit cost based on:

  • The LLM model used (pricing varies by model)
  • Token count (input tokens + output tokens)
  • Tool executions (some tools have per-call costs)

Aggregate cost data is available at the project level under billing reports.

💡

Reduce costs by using smaller models for simple tasks, caching frequent queries, and limiting max tokens in Quick Prompt nodes.

Identifying issues

Common patterns to watch for:

  • High failure rate — Check the most common error messages. Often caused by external service outages or misconfigured tools.
  • Slow execution — Look for bottleneck nodes (highest duration). Consider parallelizing independent steps.
  • High token usage — Review prompts for unnecessary verbosity. Use shorter system prompts and limit context window size.
  • Unexpected outputs — Inspect the run trace to see what each node produced. Often caused by incorrect variable wiring.

Iterating on your agent

Identify the improvement

Use run traces and error patterns to pinpoint what needs to change — a prompt tweak, a missing branch, or a tool misconfiguration.

Create a new version

Open the agent in the Agent Builder. Create a new version (Versioning) so production traffic is unaffected while you iterate.

Make changes and test

Edit the graph, test with the Test panel (Test the agent), and verify the fix.

Promote to production

Once satisfied, promote the new version. The deployment automatically picks up the new production version.

Monitor the improvement

Watch the runs list to confirm the fix is working — lower error rates, faster execution, or better outputs.

Audit logs

Project-level audit logs (accessible under Security in the sidebar) capture:

  • Agent creation, modification, and deletion events
  • Version promotions
  • Deployment changes
  • Team membership changes

Use audit logs for compliance, debugging access issues, or understanding who changed what.

Common issues

  • “No runs showing” — The agent hasn’t received any traffic yet. Send a test request via the deployment URL or use the Test panel.
  • “Runs show but no token data” — Token tracking is only available for Quick Prompt and Agent Call nodes. Pure logic flows (Condition, Loop, Code) don’t consume LLM tokens.
  • “Cost seems too high” — Check which model is configured. Larger models cost more per token. Switch to a more efficient model for simpler tasks.

Summary

You’ve completed the full Task Agent lifecycle: create → build → connect data → add tools → test → version → publish → deploy → monitor → iterate.

9. DeployOverview