This is a submission for DEV's Summer Bug Smash: Clear the Lineup powered by Sentry.

Project Overview

I built an AWS Security Posture Agent: five specialist AI agents that scan your AWS account for security misconfigurations, map findings to CIS benchmarks, score risk, and generate copy-paste fix commands.

The agents run sequentially on CrewAI with Amazon Bedrock Nova Pro as the LLM:

1. ResourceDiscovery    → inventories EC2, S3, Lambda, IAM, SGs, API GW, DynamoDB
2. SecurityScanner      → finds open ports, public buckets, admin roles, insecure configs
3. ComplianceChecker    → maps to CIS AWS Foundations Benchmark
4. RiskScorer           → severity × blast radius × exploitability
5. RemediationPlanner   → generates AWS CLI fix commands

Each agent has custom boto3 tools that make real AWS API calls against a live account with 90 IAM roles, 14 S3 buckets, 9 security groups, and 7 Lambda functions. Not test data. Real findings.

97 security findings
├── CRITICAL: open SSH/RDP from 0.0.0.0/0, IAM users without MFA
├── HIGH:     admin roles, unencrypted EBS, missing public access blocks
├── MEDIUM:   deprecated Lambda runtimes, missing versioning, stale SGs
└──

Demo

Here's the full scan running against my AWS account. The scan portion is sped up 4x, results walkthrough is at normal speed:

Bug Fix or Performance Improvement

My first instinct was to blame Bedrock latency. Every time something is slow you blame the LLM, right?

The SecurityScanner agent was taking 22.6 seconds while every other agent averaged 5-10 seconds. Without visibility into what was happening inside each agent's execution, I would have added time.time() calls and guessed.

Sentry's trace waterfall told a different story. The real problem wasn't Python execution time or network latency. It was the LLM getting a context payload it couldn't process cleanly on the first attempt.

The root cause: my IAMAnalyzer tool was fetching all 90 IAM roles from the account (59 after filtering service-linked roles), serializing them into a 27KB JSON blob, and handing that entire payload to the LLM as tool output. The context window got overwhelmed. CrewAI's internal retry logic kicked in, burning tokens on a second attempt with even more context.

One tool. Wrong default. The entire pipeline suffered.

Code

PR with the fix:

fix: paginate IAM analysis and add token budget guard #1

simplynadaf posted on Jul 20, 2026

What

The IAMAnalyzer tool was fetching every single role in the account (90 of them, 59 after filtering service-linked ones) and dumping the full details into a single JSON blob. That blob hit 26,980 characters. The LLM choked on it, CrewAI retried the task, and the SecurityScanner agent ended up taking 22.6s while every other agent finished in 5-10s.

How I found it

Added Sentry spans to each agent and tool. The trace waterfall made it obvious: SecurityScanner was twice as wide as everything else. Drilled into the tool spans and saw result_length_chars: 26980 on iam_analyzer vs ~4000 on the other tools.

7x output disproportion. That was the problem.

What I changed

Three things in iam_analyzer.py:

1. Paginate and sort roles by RoleLastUsed date, only analyze the top 20 most active ones. The rest are stale roles nobody has touched in months.
2. Skip 31 service-linked roles upfront (you cannot modify them anyway, auditing them is noise).
3. Token budget guard at the end: if the JSON output exceeds 4000 chars, trim the role_summary down to just names and policy lists.

Numbers

• Tool output: 26,980 chars down to 15,532 (42% smaller)
• SecurityScanner: 22.6s down to 17.8s (21% faster, no more LLM retry)
• API calls to IAM: 59 down to 20
• Total pipeline: 62s down to 57.7s
• Findings: still 27. No coverage loss because the top-20 most active roles contain all the problematic ones (AdminRole, Bedrock-Lambda-Role, etc are all heavily used).

View on GitHub

The core change lives in src/security_posture/tools/iam_analyzer.py. Here's the before and after:

Before (the bug):

# Fetches ALL roles without pagination limit
roles = iam.list_roles(MaxItems=100)
role_details = []

for role in roles["Roles"]:
    role_name = role["RoleName"]
    if role.get("Path", "").startswith("/aws-service-role/"):
        continue
    # Analyzes every single role...
    attached = iam.list_attached_role_policies(RoleName=role_name)
    # ...builds massive JSON output

This produces 26,980 characters of JSON for an account with 90 roles. The LLM chokes on it.

After (the fix):

# FIX: Paginate and sort by relevance
all_roles = []
paginator = iam.get_paginator("list_roles")
for page in paginator.paginate():
    all_roles.extend(page["Roles"])

# Filter service-linked roles (31 roles, can't modify anyway)
auditable_roles = [
    r for r in all_roles
    if not r.get("Path", "").startswith("/aws-service-role/")
]

# Sort by last used date (most active first)
auditable_roles.sort(key=_last_used_sort_key, reverse=True)

# Take only top 20 roles
roles_to_analyze = auditable_roles[:max_roles]

Plus a token budget guard at the end:

# Token budget guard: truncate if output exceeds threshold
if len(output) > 4000:
    result["role_summary"] = [
        {"role_name": r["role_name"], "policies": r.get("attached_policies", [])}
        for r in role_details
    ]
    result["note"] = "Role details truncated to stay within token budget"
    output = json.dumps(result, indent=2, default=str)

Three changes. Pagination, relevance sorting, and a safety valve. The SecurityScanner stopped retrying.

My Improvements

Results first:

The 21% improvement on the SecurityScanner came from the LLM completing analysis in a single pass instead of retrying.

How I found it:

The fix itself is simple. The interesting part is how Sentry pointed me straight to it.

Without the trace waterfall, all I would have seen is "pipeline takes 62 seconds." The trace said otherwise:

1. Wrap each agent execution in a gen_ai.invoke_agent span
2. Wrap each tool call in a gen_ai.execute_tool span
3. Run the pipeline and look at the trace waterfall
4. The SecurityScanner span was visually obvious: twice the width of everything else
5. Inside it, the iam_analyzer tool span showed a result_length_chars of 26,980
6. Compare to security_group_analyzer at 4,200 chars and s3_config_checker at 3,800 chars

The disproportion was the clue. If the tool output is 7x larger than its siblings, reduce it.

Best Use of Sentry

I used Sentry's AI Agent Monitoring to instrument a multi-agent CrewAI pipeline from scratch. This isn't a web app or API. It's five autonomous AI agents making LLM calls and executing custom tools. Standard APM wouldn't help here.

What I instrumented

Every pipeline run creates a Sentry transaction with this span hierarchy:

Transaction: "Security Posture Scan" (57s)
├── gen_ai.invoke_agent: ResourceDiscovery
│   └── gen_ai.execute_tool: aws_resource_scanner
├── gen_ai.invoke_agent: SecurityScanner
│   ├── gen_ai.execute_tool: security_group_analyzer
│   ├── gen_ai.execute_tool: s3_config_checker
│   ├── gen_ai.execute_tool: iam_analyzer
│   ├── gen_ai.execute_tool: ec2_security_checker
│   └── gen_ai.execute_tool: lambda_security_checker
├── gen_ai.invoke_agent: ComplianceChecker
├── gen_ai.invoke_agent: RiskScorer
└── gen_ai.invoke_agent: RemediationPlanner

The instrumentation code

For each agent (in main.py):

with start_span(
    op="gen_ai.invoke_agent",
    name=f"invoke_agent {agent_name}",
) as agent_span:
    agent_span.set_data("gen_ai.operation.name", "invoke_agent")
    agent_span.set_data("gen_ai.agent.name", agent_name)
    agent_span.set_data("gen_ai.request.model", "amazon.nova-pro-v1:0")
    agent_span.set_data("gen_ai.pipeline.name", "security-posture-scan")

    task_output = task_obj.execute_sync(
        agent=task_obj.agent,
        context=context,
        tools=task_obj.agent.tools,
    )

    agent_span.set_data("duration_seconds", round(elapsed, 2))
    agent_span.set_data("output_length_chars", len(str(task_output)))

For each tool (decorator in monitoring.py):

def trace_tool(tool_name: str):
    def decorator(func):
        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            with start_span(
                op="gen_ai.execute_tool",
                name=f"execute_tool {tool_name}",
            ) as span:
                span.set_data("gen_ai.tool.name", tool_name)
                result = func(*args, **kwargs)
                span.set_data("result_length_chars", len(result))
                try:
                    data = json.loads(result)
                    if "findings_count" in data:
                        span.set_data("findings_count", data["findings_count"])
                except (json.JSONDecodeError, KeyError):
                    pass
                return result
        return wrapper
    return decorator

What Sentry revealed

The trace waterfall made the bottleneck visually obvious. The SecurityScanner span was nearly twice the width of any other agent, 22.6 seconds while others averaged 5-10s.

Inside it, the iam_analyzer tool span showed result_length_chars: 26980 while s3_config_checker showed 3,800 and security_group_analyzer showed 4,200. The granularity goes down to individual boto3 calls. Every GetPublicAccessBlock, ListRoles, and ListAttachedRolePolicies shows up as its own span.

Before the fix (SecurityScanner dominates the trace at 22.6s):

After the fix (all agents proportional, no single bottleneck):

The iam_analyzer output dropped from 26,980 chars to 15,532, and the SecurityScanner stopped triggering retry logic.

Sentry features used

Why this matters for AI agent observability

Standard logging tells you an agent "finished." It doesn't tell you which tool inside which agent returned a 27KB payload that triggered a retry you never asked for.

With five agents each making their own LLM calls and tool executions, you need span-level visibility. Sentry's gen_ai.invoke_agent and gen_ai.execute_tool conventions gave me exactly that. I could see the problem in the trace waterfall before I even looked at the code.

That's the difference between "add some logging" and actual AI observability.

If you're running multi-agent pipelines, what's your observability setup? Curious if anyone else has hit similar tool-output-size problems with CrewAI or LangGraph.

The agent found 97 real security findings in my AWS account. The fix is in production.