closed beta · LangGraph, CrewAI & Claude Code supported · Book a Demo →
Home How It Works Frameworks Reports Claude Code FinOps FAQ GitHub Discord
Auto-stop

Watches agent traffic as it happens. Detects loops, retry storms, subagent explosions — and halts the agent before the next call fires. Nothing to configure after install.

live during run
FinOps analysis

After each task completes, agentSonar analyzes spend by pattern — caching misses, redundant calls, oversized models — and surfaces the fix for the next run.

after task completes

One report. Spend, waste, and the fix.

A concept preview of the FinOps report, with sample data. Your report shows your own sessions, your own numbers.

agentsonar · spend report · last 30 days · sample data · yours will show your own agents
total agent spend
$0
company-wide AI workflow · this month
1
saved by auto-stop
$0
Researcher ↔ Writer loop halted at 15 trips
would have kept running without agentSonar
2
FinOps savings available
$0
apply the recommendations below
wasted if ignored this month
example · caching win
$336 /day
same prompt every hour: $360/day uncached
$24/day with caching · 93% reduction
2

Where the money went

useful work · $32,000 · 80% waste · $8,000 · 20%
inside the $8,000 of waste — FinOps can recover this
caching misses
$4,200 redundant calls
$2,100 oversized models
$1,200 long outputs
$500

Cost by agent step · top 4 of 23

agent stepcallscache hitscostflag
Writer → drafting1,18012%$11,300low cache hits
Researcher → web search4,21084%$9,100·
Coder → file reads8,40071%$3,9001,240 repeats
Reviewer → Researcher312n/a$2,840silent loop

98% of calls priced from exact usage numbers · 2% estimated and marked as estimates

Fix queue · sorted by monthly impact · delivered in your report

critical · silent loop

The Reviewer keeps bouncing work back to the Researcher. 312 round trips, $2,840 this month.

✓ auto-stop already catching this
$2,400 to $3,100 /mo
compared to its last 14 days
warning · oversized model

Simple labeling tasks run on your most expensive model. A smaller model gets the same results on this task.

See the cheaper swap
$1,100 to $1,600 /mo
checked on 312 past tasks
warning · low cache hits

The Writer's prompt changes on every call, so caching never kicks in. Only 12% of its calls hit the cache, against 84% everywhere else.

Keep the prompt opening the same
$700 to $1,000 /mo
based on this month's cache misses
warning · redundant work

The same file was read 1,240 times with nothing changed in between.

Reuse the first result
$480 to $640 /mo
based on this month's repeats

Post-run. Every pattern, priced.

After each task completes, agentSonar analyzes what was spent and surfaces the exact patterns costing you money — with the fix ready for the next run. These are the ones we catch today, with more on the way.

Loops & runaway agents

Detected in real time. Auto-stop halts the agent before the next call fires — no code changes needed.

auto-stop · live

Prompts that skip the cache

A prompt that runs every hour at $15/call costs $360/day. Cached, it's $24. We flag it and show the fix.

post-run analysis

Redundant calls

The same tool called with identical inputs, over and over. We count the repeats and price the waste.

post-run analysis

Oversized models

Simple tasks running on your most expensive model. We flag the step and suggest a cheaper swap that gets the same result.

post-run analysis

Bloated outputs

Steps writing five times more than the next step reads. Shorter outputs, same quality, lower cost.

post-run analysis

More, always

Every agent codebase we instrument surfaces waste we hadn't named yet. New detections ship continuously. If you see something we don't catch, tell us.

growing

A tool-call loop cost $437 overnight in one incident. A redundant-work storm fired 14,000 identical calls. One multi-agent loop ran 11 days and cost $47,000. The failures that break agents are the same failures that burn budget.

Private preview — tell us your stack and rough monthly agent spend and we'll get you in. Email founders@agent-sonar.com →