Agent Operations
Agent work keeps running. Keep it accountable.
Loops, schedules, automations, monitors, orchestration graphs, and spawned agents can span context windows, repositories, machines, and quota windows. Ottto connects recurring work, recorded lineage, context, and cost so you can inspect what is still useful.
You decide what stays useful. Ottto provides evidence, not automatic action.
Agent operations
Illustrative data Nightly repository reviewActivity across 14 days
14 days agoToday
Inspect
One activity window crossed the next expected cadence. Open both runs to decide whether the overlap is intentional and still useful.
Schedule evidenceRun historyTokens + costCoverage
One operating view
See the work behind the spend.
Move from portfolio signals to the underlying runs, sources, and coverage before making a change.
01
Recurring groups
Review cadence, overlap, run history, tokens, and cost for work linked by supported schedule, template, title, or timing evidence.
02
Agent families
Trace recorded parent-child launches, fan-out, and descendant cost back to the root run.
03
Context posture
See what filled measured context windows, how much it occupied, and where coverage stops.
04
Prompt-cache economics
Separate cache reads, cache writes, and ordinary input before deciding whether a workflow should change.
05
Work Patterns
See when recorded agent activity clusters by local weekday and hour—without exposing prompt text.
06
Quota pace
Track provider-reported windows, resets, credits, and remaining headroom without translating capacity into savings.
Repeated work
Review recurring work as a system, not a list.
Agent Groups brings related runs together so you can compare cadence, overlap, history, tokens, cost, models, machines, and repositories in one place.
- ✓
Know why runs were grouped. Schedule, template, title, and recurring-pattern evidence stay distinct.
- ✓
Keep recorded runs in reach. Drill into history instead of trusting an aggregate alone.
- ✓
Inspect overlap. Compare activity windows, then decide whether the cadence is still useful.
Agent GroupsCadence, overlap, and history
Illustrative dataNightly repository reviewSchedule + template
Daily · 01:001418.2M$84.20 Dependency monitorRecurring pattern
About every 6h279.6M$41.08 Release verificationTitle + repository
Irregular86.1M$29.44 Review signal Nightly repository review 14 runs · $84.20
Schedule + template evidence · daily at 01:00
i
“Recurring pattern” means timing repeated. It does not identify what launched the runs.
Agent FamiliesRecorded launcher topology
IllustrativeROOTRelease coordinator$18.40 · 3.2M tokens
CHILDCode review$5.80
CHILDTest run$7.10
CHILDDocs check$2.30
CHILDVerifier$1.90
Illustrative family total$35.50Root + recorded descendants
Recorded fan-out 1 root → 4 descendants $35.50 family total
Inspect the recorded descendants before changing the root workflow.
Agent lineage
See cost across recorded agent fan-out.
One root session can launch reviewers, testers, research agents, and verifiers. Agent Families connects supported parent-child launches and rolls recorded descendant cost back to the family.
MapTimelineListDetails
Only launcher evidence establishes lineage. Similar titles, timing, or repositories are not enough; missing edges remain missing.
Context posture
See how measured context differs by repository.
Compare startup instructions, files, tool output, memory, and other measured context within one workspace. Estimated units and coverage stay visible.
Context compositionRepository A · selected range
Illustrative Startup instructions 31% Files and docs 27% Tool output 24% Memory 10% Other measured 8%
Measured context172Kestimated tokens
Largest measured share Startup instructions · 31% 172K estimated tokens
Selected range · measurement mode, source, and coverage stay attached.
Within-workspace comparisonRepositories · 30 days
IllustrativeRepositoryStartupPeakCoverage
service-api
54.5K881K42 / 47 web-console
31.2K604K36 / 40 infra
18.7K392K21 / 29 Your workspace only
Compare your own repository instructions and workflow choices. This is not a team ranking, universal benchmark, or “AI-native” score.
Highest measured peak service-api · 881K 42 of 47 runs covered
Compare only inside this workspace and selected range—not as a universal score.
Prompt-cache economics
See whether prompt caching is paying off.
Ottto separates cache reads, cache writes, and ordinary input against supported model and billing evidence. Review write cost, reuse, cadence, and cache lifetime together.
Reads Reused prompt context at the applicable cache-read rate. Writes Context stored at its applicable write duration and rate. Ordinary input Tokens billed without a cache read.
Incomplete model, pricing, or cache evidence stays incomplete—not converted into a savings estimate.
Prompt cacheExample run family
IllustrativeIllustrative input cost basis$24.80example run family
Cache reads9.4M tokens$7.10
Cache writes1.6M tokens$9.60
Ordinary input2.1M tokens$8.10
Inspect
Write cost is material beside reuse. A person should compare cadence and cache lifetime before changing the workflow.
Work PatternsAverage sessions with activity
IllustrativeMonTueWedThuFriSatSun
12a6a12p6p11p
Tue · 7–10 AMOne of the quietest windowsSun · 7 PMBusiest recurring hour
Busiest recurring hour Sunday · 7 PM Inspect schedules in this window
Average distinct sessions with recorded activity · local time.
Work patterns
See when agent work clusters.
The Work Patterns heatmap averages distinct sessions with recorded activity by local weekday and hour. Use busy and quiet windows to inspect schedules and coordinate maintenance.
- ✓
Visible denominator. Each cell shows its average and occurrence count.
- ✓
Local time. Timezone and daylight-saving boundaries are handled explicitly.
- ✓
No productivity score. Activity is an operating pattern, not a judgment about a person.
Quota pace
See quota pressure before the next run.
Put provider-reported windows, reset timing, model-scoped limits, credits, and overflow coverage beside agent activity. Headroom describes capacity—not completion, dollar spend, or savings.
Provider-reported windowsSelected account
Illustrative5-hour windowResets in 1h 42m
28% left
Weekly windowResets in 3d 9h
54% left
Review modelModel-scoped window
12% left
Inspect capacity
Review the runs and provider evidence before deciding whether to reschedule, reroute, or change a model.
Least headroom Review model · 12% left Inspect before the next run
Provider-reported capacity—not completion, dollar spend, or savings.
Evidence and coverage
Know what each signal can support.
Source, scope, freshness, measurement mode, and missing coverage stay beside the chart so estimates, partial reads, and absent relationships remain explicit.
Exact or imported
Attach the source
Provider, launcher, collector, or billing evidence stays with the fact it supports.
Estimated
Expose the basis
Estimated tokens and catalog-priced cost stay labeled and separate from imported amounts.
Partial coverage
Show the gap
Measured-run counts and unavailable fields bound the conclusions you can draw.
Not measured
Absence is not zero
Missing lineage, context, quota, or capability evidence stays unavailable—not unused or free.
Evidence boundaries
Clear signals. Clear limits.
Agent Operations supports investigation and human decisions. It does not infer intent, outcomes, or automatic action.
Does Ottto decide that recurring work is stale?
No. Ottto brings recurring work, activity, cost, context, and evidence together for a person to inspect. It does not infer staleness, stop runs, or change schedules.
Does a regular interval prove automation?
No. Repeated timing is labeled as a recurring pattern unless schedule, template, launcher, or other supported evidence establishes more.
Is repository comparison a universal benchmark?
No. It compares repositories only inside the selected workspace and time range. Ottto does not assign an “AI-native” score or rank teams against an external benchmark.
Does quota headroom mean money saved?
No. Quota windows describe capacity and pressure. Cost, subscriptions, credits, and cloud-billing evidence remain separate.
FinOps for agent operations
Make recurring agent work reviewable.
Connect supported local agent history with subscription, API, and cloud-billing evidence. Then inspect cadence, lineage, context, cost, and quota in one operating view.