Agent runs, spend & limits

Who this is for: Owner — requires core.agent.read. Where: Settings → Agent Runs

Some things in Nerve run on their own — a Protocols classification loop, a mined play candidate, a submitted job. Each is an agent run: a durable loop with a journal, a budget, and a way to stop and ask a person.

This screen is where you watch all of it. Nothing here is a black box by design — if a model did something on your property's behalf, it is on this page.

Three tabs: Runs, Attention and Spend & limits.

Note

Agent Runs appears once the agentic runtime is switched on for your organisation. Until then the surface is not there, which is the intended pre-launch state rather than a fault.

Runs#

One row per run, each with its status, the tokens it consumed and the number of tool calls it made.

Status	Meaning
Pending	Minted, not started
Running	Working
Waiting	Stopped for a person — see Attention
Succeeded	Finished with a result
Failed	Finished with an error
Halted	Stopped deliberately — a budget ceiling, a kill switch
Expired	Its deadline passed

Halted is not failed. A run that hit its token ceiling and stopped did exactly what it was told to do. It is separated out so that hitting a limit reads as a limit rather than as a bug.

An empty list reads "A run is minted when a change clears one of your salience rules. Nothing has cleared one yet." — which is a statement about your property, not about the feature being unavailable.

The journal#

Opening a run shows every step it took, in order and typed:

Step	What happened
Model	A model was asked something
Tool	A capability was called
World view	It read the state of the property
Interrupt	It stopped to ask a person
Clock	It read the time
Random	It drew a random value
Cache	A previous identical step was reused

Clock and random being journaled steps is the detail that makes the rest trustworthy. A loop that reads the wall clock or rolls a die is not reproducible unless those readings are recorded too — with them on the journal, a run can be replayed and will make the same decisions, which is what "durable" is actually claiming. Cache hits are marked, so a resumed run is visibly resuming rather than quietly redoing work you already paid for.

Attention#

Where runs stop and wait for a person, with a deadline.

Two kinds, and the difference matters:

• Confirmation — "I am about to do this, confirm."
• Approval — "This is outside what I may decide, approve it."

You Approve or Reject. An interrupt is pending, answered, escalated, or expired — and an expired one is a real outcome, not a lost message: the run stopped rather than proceeding without you. Nothing here proceeds on a timeout.

Two further panels sit on this same tab, below the interrupts:

Decisions — why the runtime chose to act, or not to. Each decision is fired, suppressed or coalesced. Coalesced is the interesting one: several triggers that would have produced near-identical work are collapsed into one run rather than each minting their own. Suppressed and coalesced entries are shown, not hidden — the times the runtime decided not to do something are as much a part of understanding its behaviour as the times it did.

Subscriptions — which event streams the runtime is following and where its cursor has reached. This is the first place to look if a loop seems to have gone quiet.

Spend & limits#

This tab is the money.

The period's consumption across the top, the plan's allowance beneath it, and your own ceilings below that — which are the ones you can actually edit.

Your plan's allowance#

Shown	What it is
Plan	The plan the allowance comes from
Included tokens	What the plan covers each month
Remaining this month	What is left of it

A plan without a metered ceiling shows No plan limit.

Usage past the included allowance appears on your invoice as a usage overage line, itemised separately from the subscription and quoting the tokens and the per-1,000 rate — so the total is always the sum of lines you can check rather than one number to take on faith.

Your own ceilings#

Separately from the plan, you can set your own limits:

• Monthly token quota
• Max tokens per run
• Max tool calls per run
• Max iterations per run

These are yours to tighten. You can lower your budget below the plan allowance; you cannot raise it above. The per-run ceilings are the ones worth setting even if you are comfortable with the monthly figure — they are what bounds a single loop that has misunderstood its task, and they turn that into a halted run rather than a month's budget.

If spend is paused#

Spend can be paused for an organisation, and when it is, this tab says so with the reason. No new runs are minted while a pause is in effect.

A pause stops future spend. It does not erase history: past runs, their journals and the charges already accrued are unchanged. Resuming clears the pause and its reason together.

Everything here is also in the audit log#

Background model calls ride the same trail as human actions — one entry per metered run, alongside every pause and resume with its actor and reason. There is no separate AI log. See Audit log.

What is recorded about a run is the model, the token counts, the cost and a digest of the ask — never the prompt itself.

What's next#

• Protocols — the main thing these runs are doing.
• MCP agents — jobs, which execute as agent runs.
• Audit log — the shared trail.