Agent Audit Log & Explainability Dashboard

The Agent Audit Log & Explainability Dashboard answers the core enterprise compliance question: "Why did the AI make this decision, what did it consider, and what did it change?"

Existing observability surfaces show what ran. This dashboard surfaces why decisions were made — every reasoning step, every tool call, every error — in a queryable, traceable UI.

Accessing the Dashboard

Global (cross-product)

Navigate to Audit Log in the account-level sidebar, or go directly to:

/dashboard/agent-audit

Product-scoped

To see audit data for a single product, navigate to that product and select Audit Log, or go to:

/dashboard/products/[id]/agent-audit

The product-scoped view pre-filters all queries to the specified product ID.

Dashboard Layout

Summary Stat Cards

Four cards at the top of the page provide fleet-level (or product-level) at-a-glance metrics:

Card	What it shows
Total Entries	Total log entries in the selected scope/time range
Tool Calls	Number of `tool_call`-level log entries
Reasoning Logs	Number of `thinking`-level log entries
Error Rate	Percentage of entries that are `error`-level

Agent Activity Distribution

A colour-coded proportional bar below the stat cards shows how log volume is distributed across agent types (e.g. Implementation, Research, Testing). Hover over a segment to see the agent name and exact count. The legend below the bar lists each agent type with its count and error count.

Three-Tab Main View

Audit Feed

A live, paginated log stream across all agents and products.

Filtering options:

Search — full-text search across log messages (matches are highlighted inline)
Level — filter by log level: thinking, tool_call, tool_result, error, metric, info
Agent type — filter to a specific agent (e.g. Implementation, Research)

Each row shows:

Level badge with icon and colour coding
Log message (with search term highlighting)
Tool name badge (for tool_call/tool_result entries)
Agent type, project name, and timestamp
Expand arrow to reveal full metadata inline
Trace button — opens the Job Trace sheet for the associated agent job

Pagination is cursor-based (ISO timestamp), so the feed remains stable as new logs arrive.

Error Audit

Filtered to error-level entries only, enriched with causal context:

Stack trace (expandable)
Pipeline trigger that caused the run
Feature title being implemented
Project the error occurred in

Use this tab to quickly identify which pipeline runs failed and trace the cause back to a specific feature or trigger.

Reasoning Chain

A chronological timeline of all thinking-level logs — the AI's inner monologue as it deliberated on a task.

Filterable by agent type
Shows the full thought text, agent type, project, and timestamp for each entry
Useful for auditing why the AI chose a particular approach or rejected an alternative

Job Trace Sheet

Clicking Trace on any audit feed row (or any entry linked to an agent job) opens a side sheet that slides in from the right. The sheet displays the complete decision trace for that single agent job, organised into four phases:

Phase	Log levels included	What it shows
Reasoning	`thinking`	Every deliberation step — why the agent was considering an action
Actions	`tool_call` + `tool_result`	Paired tool invocations with tool name, input, output, and duration
Decisions	`metric`, `info`	Conclusions reached, metrics recorded, and informational checkpoints
Errors	`error`	All errors with full stack traces

The sheet also shows:

Job summary — agent type, pipeline run ID, total duration, entry count
Files written — list of files the agent wrote during this job, with line counts
Pipeline run link — direct link to the associated pipeline run for further investigation

The sheet can be scrolled independently of the main dashboard and dismissed without losing your current feed position.

Log Levels

Level	Colour	Description
`thinking`	Purple	Agent reasoning / deliberation
`tool_call`	Blue	Agent invoking a tool
`tool_result`	Green	Result returned from a tool
`error`	Red	An error occurred
`metric`	Amber	A metric or measurement was recorded
`info`	Grey	General informational log

Supported Agent Types

The dashboard recognises and labels the following agent types:

research · design · implementation · testing · release · marketing · documentation · security · uiux · dependency · migration · branch_sync · performance · revenue · compliance · curation · billing_audit · architect · mission_alignment · onboarding

Technical Notes

No schema changes required. The dashboard reads entirely from the existing agentLogs table.
Cursor-based pagination is used throughout the feed to prevent offset drift on high-volume live data.
Server-side aggregation is used for summary stats; trace phase grouping is performed client-side on already-paginated data.
The Job Trace sheet uses a Sheet (slide-over) rather than a modal dialog so that long traces can be scrolled without blocking the main view.