Skip to content

AI Reasoning

Why AI instead of thresholds?

Traditional cost tools use hardcoded rules:

# The old way — fragile, needs constant maintenance
if resource_type == "EC2" and cpu_avg < 5.0:
    flag_idle()
elif resource_type == "RDS" and connections_avg < 1.0:
    flag_idle()
elif resource_type == "NatGateway" and bytes_out == 0:
    flag_idle()
# ... 50 more elif blocks ...

This breaks immediately when: - A resource has unusual usage patterns - A new resource type is added - Context matters (a dev database with zero connections on weekends isn't idle)

Argus takes a different approach: give the AI the raw signal data and let it reason.

What Claude receives

For each candidate resource, the AI has access to:

  • Cost — monthly USD from Cost Explorer / BigQuery
  • Metrics — 90-day averages: CPU, network bytes, request count, IOPS, connections, etc.
  • Instance size — current instance type/class injected alongside metrics (e.g. instance_type: db.r5.4xlarge, memory_mb: 3072)
  • Last activity — timestamp of the last meaningful CloudTrail / Audit Log event
  • Tags — owner, environment, team, purpose
  • Resource type and region — context for what the resource is supposed to do

How Claude decides

Claude applies qualitative reasoning that adapts to context:

EC2 at 0.001% CPU

i-0abc is an EC2 t3.large instance with:
- CPUUtilization avg: 0.0014% (90-day average)
- NetworkIn avg: 0 bytes
- NetworkOut avg: 0 bytes
- Last CloudTrail event: null (no events in 90 days)
- Tags: {} (no owner, no environment)
- Cost: $28.40/month

Assessment: This instance is effectively idle.
A t3.large with 0.001% CPU and zero network traffic
is not serving any workload. The absence of any tags
suggests it may be a forgotten resource. No CloudTrail
activity in 90 days confirms no recent human interaction.

Priority: HIGH ($28.40/month, confirmed idle)
Recommendation: Terminate. If data is needed, take an
EBS snapshot first.

RDS with zero connections — but recent activity

db-prod-replica has:
- DatabaseConnections avg: 0.0
- CPUUtilization avg: 0.3%
- Last CloudTrail event: 2026-06-01T14:22:00Z (6 days ago)
- Tags: {"Environment": "prod", "Owner": "platform-team"}
- Cost: $45.00/month

Assessment: This RDS instance shows zero connections but
was modified 6 days ago by the platform team. It may be
a standby replica, a recently migrated instance, or be
used only during specific batch windows. The owner tag
indicates active ownership.

Priority: LOW (owned resource, recently touched)
Recommendation: Tag for review — ask the platform team
if this replica is still needed.

RIGHT-SIZE: RDS over-provisioned

db-analytics-01 is an RDS db.r5.4xlarge with:
- CPUUtilization avg: 4.2% (90-day average)
- DatabaseConnections avg: 2.1
- instance_type: db.r5.4xlarge
- multi_az: true
- engine: postgres 14.7
- storage_gb: 500
- Cost: $1,240/month

Assessment: This database is severely over-provisioned.
CPU at 4.2% and fewer than 3 concurrent connections over
90 days on a db.r5.4xlarge (128 GB RAM, 16 vCPUs) indicates
the workload could run comfortably on a db.r5.xlarge.
Multi-AZ is also enabled, doubling the instance cost on a
DB that is barely used.

Priority: HIGH ($1,240/month, clear right-sizing opportunity)
Recommendation: RIGHT-SIZE: db.r5.4xlarge → db.r5.xlarge
(~$900/month savings). Also evaluate disabling Multi-AZ
(~$620/month additional savings) if this is not a
production-critical database.

Priority rules

The system prompt defines the priority framework:

Priority Condition
HIGH Confirmed idle AND costs > $20/month
MEDIUM Likely idle OR costs $5–$20/month
LOW Possibly idle OR costs < $5/month

Claude applies these rules but also exercises judgment — a $3/month resource that is definitively orphaned (no tags, no activity, no owner) might still be flagged HIGH if there are many of them in the account.

System prompt

The system prompt is built once per scan in core/agent/prompts.py:

  • Identifies the agent role and mission
  • Lists accounts and regions being scanned
  • Provides the investigation workflow (7 steps)
  • Defines what to look for (6 categories of waste)
  • Sets the priority rules
  • Enforces efficiency rules (batch cost calls, don't over-investigate)

The prompt is pinned in Anthropic's cache — subsequent loop iterations reuse the cached version, cutting the cost of the multi-turn conversation significantly.