Skip to content

Runs & Executions ​

When a customer gets a wrong answer from your support agent, the first thing the person who built it wants to know is: what exactly did the agent do? The Runs screen answers that — node by node, message by message, credit by credit — without leaving the builder.

Every time an agent starts — an inbound message, a webhook, a schedule, a manual test — Perfox records the run so you can replay it on the canvas, read it as a transcript, and see what it cost.

New to the builder? Start with Canvas Basics. This page covers what happens after a run fires.

How the screen is laid out ​

Open an agent (Build → Agents → your agent) and click Runs in the toolbar (next to Editor). The screen has three parts:

AreaWhat it shows
LeftThe list of recent runs, with a status filter and auto refresh.
CentreThe agent's canvas, read-only, coloured by how each node did in the selected run, with a bar of run details and actions on top.
BottomA resizable panel with Chat, Logs, Diagram, Tokens and Usage tabs.

The Runs screen before a run is picked — the runs list on the left and "Select an execution from the left to inspect."

Each run has its own address, so you can bookmark a run or send a colleague a link that opens it directly. Click Editor to go back to building.

The runs list ​

Runs are listed newest first.

  • Status filter — All, Pending, Success, Failed, Cancelled.
  • Auto refresh — tick it to check for new runs every few seconds; your choice is remembered.
  • Test runs you start from the builder — the Web Chat live preview, Execute or Run now — are listed here too, including runs of an active agent's unpublished changes.
  • Each row shows the start time, a status pill and how long the run took. A run that's still going shows a live, ticking timer instead. Rated runs show 👍 / 👎, and tags appear as chips.
  • Load more — 50 runs at a time; the button shows how many of the total you've loaded.
  • With no runs yet: "No runs yet. Run the workflow once to see history here."

The replayed canvas ​

Pick a run and the canvas shows the agent exactly as it is, read-only — you can select and inspect nodes, but not move or rewire them. Opening a node shows what it actually received and produced in this run (see The Properties Modal). Each node is coloured by what happened to it in this run:

Node statusColour
successgreen
errorred
paused (waiting)amber
skippedgrey

Connections show how many items passed along them, and the path the run actually took is highlighted — so a run that branched lights up the branch it followed.

The run bar ​

Above the canvas, the selected run shows its start time, a status pill, the total duration (ticking while it's still running) and a short run ID (hover for the full id). Its actions:

ActionWhat it does
👍 / 👎Rate the run good or bad for QA (click again to clear). Ratings show in the runs list.
+ Add tagAdd a free-text label (Enter to save). Tags appear as removable chips — up to 20 per run, 40 characters each.
Force end (live runs only)After a confirmation, ends the run and hangs up any live voice call (the recording and billing are finalised).
Debug in editorJumps back to the editor with this run's data loaded as the trigger's sample input, and opens the first failed node (or the last node that ran) on this agent's canvas. The loaded data is masked the same way as the rest of the run (see Masked data); if it can't be loaded, a message says so.

The bottom panel ​

Drag the top edge of the panel to make it taller or shorter — the height is remembered. ↗ Raw log (top-right) opens a plain-text log of the whole run in a new browser tab: every message, every tool call with its input, output and timing, and every node step, ready to save or share.

Runs that involved a conversation open on Chat. Webhook, schedule and manual runs have no transcript, so they open on Logs. A count on the right changes with the tab — steps, calls or messages.

Chat ​

The conversation for this run only, as colour-coded rows: the customer's messages, the agent's replies (with the tokens used and the model), and each tool call with its timing and an expandable view of its input and output. A failed tool call is marked in red. Other recorded events — such as how the agent decided which knowledge sources to use — are shown as data so nothing is hidden.

The Chat tab — the customer asks for support hours, the agent's reply with its token count and model, and a grounding event.

Logs ​

One row per node that ran, in order, with its status and time taken. Condition nodes show which branch was taken; a stage finishing shows a green Stage completed pill; a node the agent is waiting on shows open.

Expand a row to see:

  • Input · variables entering — what the node received;
  • Output · variables this node added / changed — what it produced, marked + added, ~ changed or - removed;
  • AI trace (AI Agent nodes) — the prompt, the reply and every tool call for that turn;
  • Raw record — the underlying data, for deep debugging;
  • an error banner, if the node failed.

The Logs tab — the trigger and AI Agent steps, with the AI Agent expanded to show its input, the variables it changed, and the AI trace.

Diagram ​

The run drawn as a sequence diagram with three lanes — User, AI Agent and Tools / MCP — and an arrow for every message, tool call and tool result (with timing). A banner on top says ✓ Completed · N steps, or ✗ Failed at step K with the tool and error — a one-glance picture of who called what, and where it broke.

The Diagram tab — the customer's question and the agent's reply between the User and AI Agent lanes, under a "Completed · 2 steps" banner.

Tokens ​

What the run used, per service and model, from the same metered usage your bill is calculated from — so voice and background services are included, not just the agent's replies.

ColumnMeaning
CallsHow many calls that service/model made.
Service / ModelThe job (e.g. the agent's main reply, or a background check) and the model that did it.
InputTokens sent to the model. Large numbers usually mean a long conversation history.
Cached (green)Shown only when some input was served from the model's cache at a much lower price — already included in Input.
OutputTokens the model produced.
Audio in / Audio out (purple)Shown only for voice — a spoken reply is billed as audio, so text output stays near zero.
TotalInput + Output.

A total row sums every column, and Credit burn (this run · all services) shows what the run cost, in credits and rupees.

System prompt · per agent turn. For runs with text replies, a block above the table shows how much of each turn's input was your agent's instructions: System prompt ≈9,145 tok + dynamic context ≈6,817 of 37,351 input (43%). The system prompt and dynamic context (per-turn context such as customer details) are estimates at about four characters per token, from the prompt that turn actually sent; input is the model provider's own count, which also includes the conversation history and tool definitions. It works for past runs too. Voice calls don't record this breakdown, so a voice run has no block.

The Tokens tab — the system-prompt share per agent turn, one row per service and model with input, cached, output and total tokens, the totals row, and the credit burn for the run.

Usage ​

A billing-style summary of the single run.

TileMeaning
MessagesCustomer + agent messages.
Metered callsEvery billed call in the run, across services.
TokensTotal tokens, split into in and out.
Tool callsIntegration and built-in tool calls.
Failed toolsHighlighted when above zero.
DurationHow long the run took.
Credit burnWhat the run cost, in credits and rupees, across all services.

Where the time went splits the run's time into LLM, Tools · MCP and Platform on a bar, and lists the slowest tools, so a slow run is blamed on the right thing. Older runs and voice calls that don't have per-turn model timing show a two-way split instead, with a note.

The Usage tab: the run's usage tiles — messages, metered calls, tokens and tool calls.

Run statuses ​

StatusMeaningColour
SuccessThe run finished. An AI Agent that replied and ended its turn counts as success.green
Completed with errorsThe run finished, but a step reported an error along the way — for example an HTTP Request that got an error response, or a data step that failed. Such steps don't stop the flow, so the run carries on; open Logs to see which step. Filter with Success to find these.amber
FailedSomething went wrong — open Logs or Diagram to see what.red
CancelledThe run was stopped (for example with Force end).grey
PendingStill running, or waiting — for a person, a reply, or a later step.amber

The same colours are used in the runs list, the run bar and on the Conversations page. One conversation can span several runs, so a customer's full history is on Conversations.

Alerts for runs nobody is watching ​

Runs started by a Webhook, Schedule, Internal Event, Knowledge Base update or Folder Watch trigger have no customer on the other end to notice a problem. When one of them fails or completes with errors, you get an in-app notification — Workflow "Lead qualification" failed or … ran with errors — naming the step and the error, with a link to the agent. Runs from channel messages, Manual Test, Sub-Workflow calls and tests you start in the builder don't alert.

  • At most one alert a day per agent, and never more than one an hour if it keeps flapping.
  • A clean run resets this, so the next failure after a recovery alerts straight away.
  • If a step's error response is expected (an HTTP step's "not found", say), add that status to the step's accepted statuses so it stops counting as an error.

Masked data ​

What you see in a run is masked the same way as the conversation transcript:

  • Secrets — API keys, authorization headers, a node's own secrets — always show as [redacted].
  • Personal data — phone numbers, emails and similar — is masked (for example +91XXXXXX2401, a***@example.com) while your workspace's PII redaction setting is on, which it is by default.

This covers every node's input, output, variable changes and error text in Logs and in the node panels, and the run data Debug in editor loads — older runs included. The run itself keeps the real values it needs to continue. If you replay a test whose sample payload still contains masked values, the test is refused and names the fields to replace with real test values.

Worked example — debugging a failed run ​

Setup. Acme Diagnostics has an agent that looks up a patient's latest report and replies by email. Priya, a customer, wrote in and got no reply.

Action. Asha opens Runs, filters to Failed, and clicks the newest red row. The canvas lights up: the trigger is green and the AI Agent is red. In Diagram the banner reads "✗ Failed at step 4 — get_latest_report: connection timeout." She clicks Debug in editor; the builder opens with Priya's run data loaded and the failing part selected.

Result. The integration's address had a typo from the last edit. Asha fixes it, sends a test through, and follows up with Priya.

What just happened. The coloured canvas showed where the run went; the Diagram pinpointed the failing tool in one line; and Debug in editor let Asha reproduce the failure with the real data instead of guessing.

Common tasks ​

Find out why a run failed. Filter to Failed, open the run, read the Diagram banner or expand the red row in Logs, then click Debug in editor.

See what a run cost. Open Tokens for the breakdown and the credit burn, or Usage for the summary and where the time went.

Rate and tag runs for QA. Use 👍 / 👎 and + Add tag in the run bar; both show in the runs list.

Stop a stuck live run. Open it (it shows Pending and a ticking timer) and click Force end.

Share or archive a run. Click ↗ Raw log and save the page, or copy the run's address.

See also ​