Observability & Logs

Overview

Observability provides a detailed execution trace for every AI agent call that occurs inside a TotalApp workflow. Unlike the Monitoring screen (which tracks workflow-level status), Observability drills into the LLM layer — showing you exactly what the model received, what it decided to do, and what it produced.

This is the primary debugging and auditing tool for AI-powered pipelines. When an agent behaves unexpectedly, Observability lets you inspect the full context window the model saw, every tool call it made, and the exact token counts billed to your API account.

Trace View

Each agent execution produces one trace entry. The trace expands into a timeline of events:

• System prompt — the full system prompt sent to the model, including any persona blocks and injected knowledge-base chunks.
• User message — the input the agent received from the upstream workflow node.
• Tool calls — each tool the model invoked, with the exact arguments it passed and the value returned by the tool implementation.
• Assistant response — the model's final text output after all tool calls resolved.
• Token usage — input tokens, output tokens, and total tokens. For cloud models, this maps directly to API billing.
• Latency — time from first token sent to last token received, broken down by model time vs. tool execution time.

Filters and Search

• Agent filter — show traces for a specific named agent or all agents.
• Time range — last hour, last 24 hours, last 7 days, or a custom date range.
• Status filter — success, error, or tool-call-only (traces where the model called at least one tool).
• User filter — in multi-user deployments, filter traces by the user whose workflow triggered the agent call.
• Search — full-text search across prompt content and agent responses to find traces containing a specific topic or error message.

Aggregated Metrics

The top of the Observability screen shows aggregate stats for the selected time range and filters:

• Total calls — number of agent executions.
• Total tokens — combined input + output tokens across all calls.
• Average tokens per call — useful for spotting prompt-size regressions after a system prompt change.
• Error rate — percentage of calls that resulted in an error (model error, tool failure, or context-length exceeded).
• Tool call rate — percentage of calls where the model invoked at least one tool. A sudden drop might indicate the agent is no longer reliably using its tools.