Runs and Activity

Runs is the complete history of everything your agents did: what started each run, how long it worked, what it cost, what it produced and the evidence behind every number. From an agent's Agent Studio page you can also start a run yourself and watch it work, step by step.

What this screen is for

A run is the unit of work in AgenticObjects. Its record is written while the run happens, not reconstructed afterwards, so it tells you exactly what the agent did and under which settings. That makes runs comparable: open two, read their effective settings side by side, and every difference is there with its source.

The record is also your assurance: a run that stops at a budget names it, waiting time is shown apart from working time, and every query is listed with where its time went.

Note: Completed steps are never repeated. If a run is interrupted and resumed, it continues from the first step that had not finished, so nothing already done is done — or paid for — twice.

What you see

1. Runs — the page title, with Every run, honestly — including cut-offs and errors beneath it.
2. The filter bar — Agent, Trigger, Status and Experiment selectors, From and To dates, Filter and Clear.
3. The run table — Time, Agent, Trigger, Status, Duration and Cost (USD). Experiment runs carry an Experiment badge, and a ⚑ flag marks a run with an open escalation. The time opens the run.
4. The run header — the run's identifier, its status, the agent and the trigger, and an Experiment badge where it applies.
5. Effective settings (frozen at run time) — a Setting / Value / Source table of the model settings and budgets the run executed under.
6. The time line — active time, suspended time and cost, followed by any cut-off notice and any Suspended episode with its Resumed after interruption at line.
7. Outputs — every finding, recommendation or brief the run wrote, with its stage.
8. Activity — the run's steps, drawn from the stored record. Below it, Step breakdown with Load step detail, and Queries: each attempt with its outcome, its rows and the time spent Waiting for a turn, Waiting for a connection, Running at the data source and Reading the rows.
9. Operation trace — What this run did: turns worked, queries sent, outputs published, cost and tokens. Beside it, An account of the agent's work with Narrative on demand.
10. Executed query (layer 3) — the statement itself for those who hold the permission; for everyone else a locked card with Request Technical Review. Below it, any Escalations, each with Mark reviewed.

How to …

Start a run by hand

1. Open the agent in the Agent Studio and go to Schedules & Distribution, then Run Now.
2. Check Last run, choose the Run kind — Answer (today's behaviour) or Discovery run — and, if you want something other than the agent's default, This run's output.
3. Press Run Now. The live Activity stream shows each step as it happens.
4. When the run finishes, a summary shows its status with a Reasoning / Evidence Chain link into the full record.

Give a single run a smaller budget

1. In Run Now, open Adjust this run.
2. Lower any of the limits shown. Anyone may LOWER budgets for this run. The limits are:
   • Step cap — how many steps the agent may take.
   • Time ceiling (s) — how long the run may take, in seconds.
   • Money ceiling (USD) — how much the run may spend on the language model.
   • Token cap — the most the model may write in one call, in tokens.
   • Most calls to the same tool — how many times the agent may call one tool.
   • Most rows shown to the model — how many rows of a query result the model reads.
3. Start the run. The lower values apply to this run only and appear in its effective settings.

Run an experiment above the normal budget

1. In Adjust this run, switch on Experiment mode. It is available to people who hold the experiment permission.
2. Raise the budgets you need, within the platform ceilings.
3. Start the run. It is recorded, audited and badged Experiment everywhere it appears.

Compare two runs

1. Filter the list to one agent and open the two runs in turn.
2. Read Effective settings (frozen at run time) on each: the Source column shows where each value came from.
3. Compare the time lines and the Queries card to see whether a difference came from the work itself or from waiting on the data source.

How to read it

• The trigger says where a run came from: Scheduled, Chat or Manual.
• Cut off is a status, not a failure. A Cut off (budget) run stopped at a named budget, and findings up to the cut are preserved below.
• The Source column turns a number into a decision. It tells you whether a value was inherited, set on the agent, lowered for this run or raised for an experiment.
• Active and suspended time are separate. A slow run and a run that waited look different, so the duration always means working time.
• The narrative is optional and kept. Narrative on demand asks the model to describe the run in a few sentences; it never sees results, so it writes no figures. It is generated once, and later readers see it at no cost.

Good to know

• Run records are permanent and never edited, so the history always shows what really happened.
• Manual runs use the scheduled configuration and never ask a clarification question — nobody is waiting to answer one.
• An experiment changes what a run may spend, never what it may reach: tools, query mode, row-level security and metric scope stay exactly as they are.
• Experiment runs record their output but never deliver it, and they are kept out of product measurements.
• The Activity stream never contains the executed query. The statement lives in its own layer, behind its own permission.

Related

• Agent Studio — where the settings a run freezes are set
• Chat — where interactive runs answer
• Spend — what runs cost across the workspace
• Trust tiers — the badge on the numbers a run produced
• How to read a brief — the delivered output of a scheduled run