02:14:11Z · run r_9f2c1a4e · agent refund-triage-v3

Find the call that cost you four hundred dollars.

AgentOps Monitor replays every tool call, every LLM decision, every dollar an agent spends, so an on-call engineer at 2 a.m. can stop guessing and start fixing.

Install the SDK Install to replay

Three lines. The planned SDK adapters target LangChain, CrewAI, and the OpenAI SDK. Concept surfaces shown below are built from the proposed data model.

Concept · concept surface, real data shape

trace r_9f2c1a4e budget exceeded

started 2026-08-14 02:14:11Z duration 48,219 ms spans 412 cost $18.44 tokens 412,880 / 88,104

+0 msrefund-triage-v3start
+12 msllmopenai.chat · 8,204 tok
+612 mstoolvectorstore.query
+820 mstoolstripe.refunds.create · HELD
+5,402 mstoolretries × 3
+11,008 msllmopenai.chat · 64,118 tok
+18,402 mstoolslack.postMessage
+48,219 msendbudget exceeded

Refund triage, three retries, one held refund.

Thesis

An agent without a budget is an incident waiting to be invoiced. Three minutes of looping, and the on-call sees the bill before they see the fix.

Four pains, in the buyer's words

01
You cannot see what it did. Did it hallucinate, call the wrong API, retry forever?
02
You cannot explain a decision. Why was that refund rejected, that lead routed there?
03
Costs spiral without a ceiling. One looping agent, one very large bill.
04
Compliance blocks the deploy. No audit trail, no approval, no launch.

Trace replay, step by step

Every tool call, every LLM decision, with the exact tokens and wall time. Click a span, see its prompt, the model output, and what it changed in the world.

Concept · concept surface, real data shape

span s_011d · openai.chat +11,008 ms

modelgpt-5-mini

tokens in54,118

tokens out10,000

cost$8.71

retry0 of 3

system: You are a refund triage agent for order #A-9123...
user: refund requested, reason: damaged on arrival
tool: vectorstore.query → 4 matches
assistant: Recommend HOLD pending review.
tool: stripe.refunds.create → status: held
tool: slack.postMessage → channel: #refunds

Span detail, prompt and tool calls.

Concept · concept surface, real data shape

timeline · 412 spans 1 hot, 1 loop

010s20s30s40s48s

Run timeline. One span runs hotter than the rest.

Spend you can defend in a budget meeting

Hard caps at the workflow, the agent, and the user. A run stops at the limit you set, not at the limit your CFO finds out about.

Concept · concept surface, real data shape

cost · this week under budget

agent	runs	cost	cap	status
refund-triage-v3	12,402	$182.04	$300	ok
support-routing-v1	8,114	$64.21	$200	ok
lead-research-v2	3,118	$48.12	$120	ok
doc-summarise-v1	22,408	$412.18	$500	82%
qa-test-gen-v0	4,802	$112.40	$200	ok

Cost per agent, hard cap, current status.

Concept · concept surface, real data shape

policy · guardrails require_approval

policy:
  refund:
    cap_per_run_usd: 50
    require_approval_above: 250
    approver_role: finance-lead
  tool:
    stripe.refunds.create:
      hold_above_usd: 100
    slack.postMessage:
      block_channel: ["#exec-private"]
  loop:
    max_retries: 3
    halt_after_seconds: 60

Policies ship in the SDK, run in our edge.

Install the SDK

Three lines. The pre-launch SDK is designed for LangChain, CrewAI, and the OpenAI SDK. Compatibility remains unverified until the package ships.

from agentops_monitor import monitor
monitor.init(api_key="aom_...", budget_usd=50)
# planned adapters: LangChain, CrewAI, OpenAI SDK

Install the SDK

Pre-launch interface. Package name and adapter compatibility are provisional.

Integrations

The names below are planned adapters, not shipping integrations or partnerships.

LangChainplanned
CrewAIplanned
OpenAI SDKplanned
Anthropic SDKplanned
Stripeplanned
Slackplanned

Pricing

Planned launch pricing for free, team, enterprise, and metered overage.

Free

$0a month

10,000 agent runs a month
Basic tracing
Community support

Install the SDK

Team

$299a month

500,000 runs
Cost governance, alerts
Slack support

Install the SDK

Enterprise

From $2,000a month

Custom run limits
SSO, audit export, custom policies
Named support

Install the SDK

Overage

Meteredper 1,000 runs

Billed monthly on actual usage
Soft cap before hard cap
Same SLA as the plan it sits on

Install the SDK

What an audit reader actually gets

The audit trail will hold every prompt, model output, tool call, and human approval, retained according to configured policy. SOC 2, HIPAA, GDPR, and ISO badges are not claimed on this page. The audit trail is the product.

event	recorded	retained
prompt	hash + length	policy window
model output	hash + length	policy window
tool call	args, response, latency	policy window
human approval	approver, role, time	policy window
policy block	rule, value, decision	policy window

You cannot ship agents you cannot see.

The bill is the first thing that tells you.

Install the SDK