Overview

Performance Insights turns what your AOPs actually do into live business reports — without anyone writing SQL, defining a table, or instrumenting a single run. From one screen you see which AOPs are reporting, start reporting on the ones that aren't, accept new and updated reports, and open any AOP's report to read the business outcomes it's delivering.

Your AI Colleagues already do work and leave a detailed trace of everything they did — every tool call, every hand-off to a helper agent, every loop over a list, every approval they paused for. Performance Insights reads those traces and builds a report out of them, deciding for itself what the business facts are, what a "row" should mean, and which charts are worth drawing. Nobody writes a query. Nobody defines a schema.

Performance Insights appears in two places:

The Insights tab on an AOP — where you configure a report, accept the proposed one, and then read it.
The Performance insights screen (Insights → Performance insights) — the workspace-wide view of every AOP and the state of its report.

The workspace screen groups every AOP by where its report sits in the reporting lifecycle:

Live — performance insights generated and capturing from every run.
Action needed — a proposed report waiting for you to accept it, either a first one or one proposed after the AOP changed.
Ready for reporting — AOPs you could report on, including ones still accumulating runs or being analysed.
Paused — reporting has stopped, either because someone turned it off or because the AOP changed substantially and the report is rebuilding.

Live and Paused tell you what's running and what's stopped; Action needed and Ready for reporting are your two queues. Together they turn "is my reporting healthy?" into a glance instead of a hunt.

Key Components

Report Lifecycle Overview — a single, searchable view of every AOP in the workspace, grouped by reporting stage.
Automatic Report Design — point Performance Insights at an AOP and it reads that AOP's past runs, then proposes the columns to capture and the charts to draw.
Continuous Capture — once accepted, every completed run is automatically turned into structured rows, with no configuration per run.
Reports & Charts — KPI tiles and trend, bar, donut and funnel charts, plus the captured rows, filterable by date range and drillable down to the rows behind any number.
Built-in Reliability Cards — every report automatically carries a completion rate and a breakdown of why runs didn't complete, so reliability means the same thing on every AOP.
Insights Agent — ask a question about a report in plain English and get an answer computed on the spot, including which runs failed and why.
Change Management — when you edit and republish an AOP, the report is kept matched to the process it describes.
Governance & Auditability — versioned report generations, an audit trail, design-time PII guards, export, and tenant-scoped retention.

How a Report Is Built

A report is built on a short assembly line. You interact with only the two ends of it — you accept the design, and you read the report.

Design. Performance Insights reads the AOP's skills, its helper AOPs, and its real past runs, then proposes what to report on, the columns to capture, and the charts to draw — writing the query behind each chart. You review the proposal and accept it. Nothing is captured or shown until you do.

Capture. Every run that reaches a finished state is captured automatically: the system reads that run and fills in the columns the design defined. The report fills as runs happen — you don't trigger or schedule anything.

Keeping current. When you edit and republish an AOP, Performance Insights sizes the change: a cosmetic edit changes nothing, a minor edit updates the report in place, and a substantial edit rebuilds the report from runs of the new version and asks you to accept it — while the previous version keeps showing until you do.

Note: Performance Insights only analyses an AOP once it has accumulated enough real runs to learn from. See Setting up Performance Insights for the run requirement.

What One Row Means (Reporting Grain)

The single most important idea in Performance Insights is grain — what one row represents — because it's what keeps every number honest.

Entity grain — the AOP processes a list of things, so there's one row per thing (one row per invoice). Counting means counting distinct things, not counting how many times an agent mentioned them.
Run grain — the AOP produces one summary each time it runs, so there's one row per run. There is no "thing"; the run is the unit.

Performance Insights infers the right grain when it designs the report, and every count and total is computed against it. That is why one business thing touched by several agents at several moments is still counted once. Grain also decides what a helper AOP contributes: a helper doesn't get a report of its own, because its data is captured into its parent's report.

Auditability & Governance

Versioned reports — a new version is created only when a substantial change is accepted. The live version shows by default and frozen versions stay viewable, so older reports remain reproducible.
Audit trail — configuring, accepting, editing, pausing and regenerating a report are all recorded as audit events.
Sensitive data kept out by design — the analysis is instructed never to propose a column or insight that would hold personally identifiable information, so PII is excluded when the report is designed rather than filtered out afterwards.
Export and retention — report data can be exported for sharing or downstream analysis, and captured rows are subject to tenant-scoped retention.

Why Performance Insights Is Critical for Business Automation

Visibility without instrumentation. Teams get a business-level view of what their automations are doing — amounts processed, approval rates, exceptions flagged — without anyone building a data pipeline, writing SQL, or maintaining a schema. The report designs itself from work that's already happening.
Outcomes, not just activity. Run history tells you an AOP ran; Performance Insights tells you what it achieved, turning execution traces into the metrics a business owner actually asks for.
Failures are first-class. Every report describes the whole picture rather than only the happy path, and the specific failing runs are one question away.
Trustworthy numbers. Deciding grain up front, refusing to design a report on insufficient evidence, and answering "how is this computed?" from the report's own stored definition produce numbers that hold up to scrutiny.
Reports that keep pace with the process. Automations evolve constantly. Because the report rebuilds itself when an AOP changes meaningfully — while never showing blank or mismatched data in between — reports stay accurate without manual upkeep.
Self-serve answers. Business users interrogate their own data in plain English, removing the dependency on an analyst for every new question.

Next: see Setting up Performance Insights for the full walkthrough, and Managing reports for editing, pausing, and versioning an existing report.