Monitoring

Monitoring covers alert rules, rate limits, and behavioral metrics for your agents, including drift and anomaly detectors, and is available on all plans including Free. Each alert reads off the signed record Asqav writes for the action, so when a rule trips you can still verify what actually happened well after the fact.

Alert Rules

Create rules that trigger alerts when agent metrics exceed thresholds:

python

import asqav

asqav.init(api_key="sk_...")

# Create an alert rule
rule = Asqav.Monitoring.create_rule(
    name="High error rate",
    metric="error_rate",
    condition="gt",
    threshold=5.0,
    window="5min"
)

print(f"Rule ID: {rule.id}")
print(f"Status: {rule.status}")

# Create a latency alert
latency_rule = Asqav.Monitoring.create_rule(
    name="p95 latency spike",
    metric="latency_p95",
    condition="gt",
    threshold=2000,
    window="5min"
)

Available Metrics

Metrics you can monitor with alert rules:

Metric	Description	Unit
`error_rate`	Percentage of failed actions	%
`latency_avg`	Average action latency	ms
`latency_p95`	95th percentile latency	ms
`latency_p99`	99th percentile latency	ms
`actions_per_hour`	Action throughput rate	count/hr
`cost_per_hour`	Hourly cost accumulation	$
`total_actions`	Cumulative action count	count

Conditions

Comparison operators for alert rule thresholds:

Condition	Description
`gt`	Greater than
`lt`	Less than
`gte`	Greater than or equal
`lte`	Less than or equal
`eq`	Equal to
`ne`	Not equal to

Rate Limits

Set per-agent rate limits to prevent runaway behavior:

python

# Set rate limit for an agent
limit = Asqav.Monitoring.set_rate_limit(
    agent_id="agent_abc123",
    max_actions=1000,
    window="1hr"
)

print(f"Limit: {limit.max_actions} per {limit.window}")

# Check current usage against limit
usage = Asqav.Monitoring.get_rate_usage(
    agent_id="agent_abc123"
)

print(f"Used: {usage.current}/{usage.limit}")
print(f"Resets at: {usage.resets_at}")

Alert Cooldown

After an alert fires, it enters a 5-minute cooldown period. During cooldown, the same rule will not fire again even if the condition remains true. This prevents alert storms during sustained incidents.