1. Docs
  2. AI
  3. Autonomous Agents
  4. How the numbers work

How the numbers work

Every number and colour on the Autonomous Agents page, how it is calculated, and an example you can check against your own project.

4 min read Updated CustomersAdmins

In short

On this page you can

  • Check what Active agents, Runs today, Proposals to review and Spent today count.
  • Work out why a run stopped at its step or cost limit.
  • Predict when the daily budget blocks new runs.
  • Read the colours of the budget meter and the height of the run bars.

How to use it

  1. Find the number on the page and jump to its section here.
  2. Compare it with the run detail (Steps, Tokens, Cost).
  3. Check the agent’s Limits and time zone in its settings.
  4. Raise a limit only after you know why it was reached.

The four cards at the top#

They appear once the project has at least one agent. They are worked out in your browser from the agents and the latest 100 runs of the project.

KPI cardsShown in: Autonomous Agents, above the tabs
Active agents        = agents switched on  /  all agents
Runs today           = runs (of the latest 100) whose start is today (your computer's date)
Proposals to review  = sum of "Proposals" of today's runs        amber when > 0
Spent today          = sum over agents of today's spend           4 decimals

Example: three agents, two switched on: 2 / 3. Today Daily triage ran once (0 proposals, $0.0123) and Library maintenance once (4 proposals, $0.2410). Runs today 2, Proposals to review 4, Spent today $0.2533.

Read it right: Proposals to review counts every proposal recorded today, also the ones you already applied or dismissed. Open the runs to see which are still waiting. “Today” for Runs today follows your computer’s date; Spent today follows each agent’s own time zone.

Run cost and today’s spend#

Cost of a runShown in: Runs › Cost (USD), run detail, card bars
model call cost = input tokens  ÷ 1,000,000 × input rate of the model
                + output tokens ÷ 1,000,000 × output rate of the model
run cost        = sum of its model call costs              (stored to 6 decimals, shown with 4)
Tokens          = input tokens + output tokens of all calls
today's spend   = sum of the costs of the agent's runs started since midnight
                  in the agent's time zone

Rates per model are set by SmartestQA; a model that is not listed uses the default rate. Tool calls cost nothing; only model decisions do. A run in Claude Desktop records no cost in SmartestQA.

Example: a run makes 6 model calls. The row shows Tokens 48,200. The step rows show costs of $0.00410, $0.00395, $0.00402, $0.00388, $0.00415 and $0.00290: the run costs $0.0230.

When a run stops#

Limit checksChecked before every model decision
at start:          today's spend ≥ Daily budget          → Budget blocked (run never starts)
before each step:  run cost ≥ Max cost per run          → Stopped at limit
                   today's spend ≥ Daily budget          → Budget blocked
after the last allowed step without a final answer       → Stopped at limit
steps allowed     = Max steps per run (1–60), one per model decision

Because the check happens before a step, a run can end a little above its limit: the last step’s cost is added after the check.

Example: Max cost per run $0.20, Daily budget $0.60. In run 1, before step 7 the cost is already $0.2050 ≥ $0.20, so it ends Stopped at limit at $0.2050. Runs 2 and 3 cost $0.19 and $0.21: today’s spend is $0.6050. Run 4 is recorded as Budget blocked without starting. At midnight in the agent’s time zone the spend starts again from $0.

Budget meter on the card#

TodayShown in: each agent card
fill %  = min(100, today's spend ÷ Daily budget × 100)
colour  = red at 100%  ·  amber from 80%  ·  grey below 80%
label   = $spend (4 decimals) / $budget (2 decimals)
a spend above 0 always shows at least a 3% sliver

Example: spend $0.4900, budget $0.60: 81.7%, so the bar is amber and reads “$0.4900 / $0.60”. One more run of about $0.11 fills it and turns it red; the next run is Budget blocked.

Last runs bars#

Last runsShown in: each agent card
bars     = the agent's latest 14 runs, oldest on the left; empty slots are short grey bars
height   = 8 px + round(run cost ÷ highest cost among the bars × 22) px
colour   = result: green Succeeded · red Failed · amber Stopped at limit / Budget blocked
           · blue Running · grey Queued
hollow   = Failed, Stopped at limit, Budget blocked     pulsing = Queued or Running

Example: the bars cost $0.02, $0.04 and $0.08. The highest is $0.08: heights 8 + round(0.25 × 22) = 14 px, 8 + 11 = 19 px and 30 px. A Budget blocked run costs $0 and is an 8 px hollow amber bar.

Duration and dates#

Formats
Duration  = finished − started (still running: now − started)
            under 60 s → "42s"      otherwise → "3m 5s"
Started   = day.month hour:minute in your browser's locale
Ago       = "just now" · "n min ago" · "n h ago" · "n d ago"

The overdue warning#

Expected run did not arriveShown in: card of a Claude Desktop agent
shown when ALL of these are true
  Runs on = Claude Desktop, agent switched on, trigger = Schedule (not After each deploy)
  now ≥ last scheduled time + 90 minutes
  the agent existed before that scheduled time
  Claude Desktop has not asked for this agent since that scheduled time

Example: Library maintenance runs every Friday at 07:00 in Claude Desktop. The laptop is closed all Friday morning. From 08:30 the card says “Expected run did not arrive (…07:00) — is Claude Desktop open?”. Once Claude Desktop asks again, the warning goes away, even if SmartestQA answers SKIP because the late cutoff has passed.

When a run is closed as stuck#

Checked every minute and at server start
Queued  with no progress for 15 minutes             → Failed
Running with no progress for 10 minutes             → Failed
Running in Claude Desktop with no result for 6 hours → Failed

This frees the agent, which can have only one run going at a time.

Didn’t find your answer?Write to support with your project name and what you see on screen.

info@smartestqa.ai
© 2026 Selenity Technology · SmartestQAsmartestqa.aiThis page as MarkdownChecked against smarttest-fe predev 2a8c79cc (AiAgentsPage, AiAgentCard, aiAgentUi), smarttest-be 633c2a03 (AiAgentRunner, AiPricingService, AiAgentLlmExecutorService)