AI Agent Operations

A live look at every running task,
with a wall-clock budget on each one

Agent Runtime is the live operator view of every task currently running in the fleet. Each task has a wall-clock budget (default 5 minutes); tasks over budget are flagged and can be stopped one-click. Use it to spot runaway tasks, cancel them, and learn what classes of work need a longer budget.

Get Started Talk to Sales
app.novaaiops.com / agent-runtime
● LIVE
5 min
Default budget
Per-class
tunable
One-click
cancel any task
Live
updates every 2s
Wall-Clock Budgets

A budget per task class, enforced

Each task class has a budget. Quick classes (route alerts) have 60s; medium classes (vacuum, scale) have 5m; long classes (large migrations) have 30m. Tasks over budget are visually flagged and can be cancelled by an operator. The platform also auto-cancels at 2x budget unless the task class explicitly opts out.

  • ✓
    Per-class budget: tuned per task class; default 5m for new classes; tunable in the runtime config
  • ✓
    Visual flagging at budget: tasks over budget shown in yellow; the operator can decide whether to keep waiting
  • ✓
    Auto-cancel at 2x: unless a class opts out, runaway tasks are cancelled at 2x budget, caps the worst case
app.novaaiops.com / agent-runtime · budgets
Cancel Semantics

Cancellation never leaves half-done work

When you cancel a task, the runtime first attempts a graceful stop: in-flight tool calls finish, partial state rolls back to the last checkpoint. If the task does not respond to graceful stop within 10s, the runtime hard-stops it. Either way, no half-applied changes survive a cancel. The cancel reason and operator id land in Agent Ledger.

  • ✓
    Graceful first: in-flight tool calls finish; partial state rolls back to checkpoint; tasks usually stop in seconds
  • ✓
    Hard-stop fallback: unresponsive tasks get hard-stopped at 10s, but checkpointing means no orphan state
  • ✓
    Logged with reason: every cancel records who cancelled, when, why, visible in Agent Ledger
app.novaaiops.com / agent-runtime · cancel
Resource Footprint

Tokens, time, and tools per task

Each running task shows live token usage, wall-clock duration, and tools called so far. The data lets operators decide whether a task that is taking too long is doing real work (lots of tool calls, lots of tokens) or stuck (no progress in 30s). Same data feeds Cost Circuit Breaker for spend attribution.

  • ✓
    Live token count: tokens consumed so far this task; lets you spot expensive runaways
  • ✓
    Live tool count: tools called so far; helps distinguish "doing work" from "stuck thinking"
  • ✓
    Per-task ledger preview: the partial ledger entry is visible during the run so you can see the agent's reasoning
app.novaaiops.com / agent-runtime · resources
Tuning Insights

Frequent over-budget = tune the class

When a class is consistently over budget, that is a tuning signal. The page reports per-class p95 duration and flags classes that exceed their budget more than 5% of the time. Either the budget is too tight or the class is running expensive work that should be split.

  • ✓
    Per-class p95: p95 task duration per class with comparison to budget
  • ✓
    Over-budget rate: over-budget per class as a percent, tuning trigger at 5%+
  • ✓
    Suggested budget: recommended budget based on actual p95 + 20% headroom
app.novaaiops.com / agent-runtime · tuning
Video walkthrough coming soon

Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.

See it, stop it, learn from it

Runaway tasks waste tokens and risk side-effects. The runtime page makes them visible and stoppable.

Get Started Request a Demo