Overview

Three coding-agent CLIs ran 10 live AWS operations tasks, with and without a task runbook, three repetitions each — 180 graded runs.

10 tasks3 agents2 arms180 runs

Tasks included

  • Investigate who made an unauthorized security-group change
  • Identify and release unattached Elastic IP addresses
  • Investigate and fix Lambda invocation errors
  • Investigate and reconcile terraform configuration drift
See all 10 tasks

Agent profiles

Cursor — best overall
98%$0.20/runsuccess
Fastest, cheapest, and leanest on nearly every task, and it pairs that efficiency with…
Codex — best at adherence
93%adherence
The only agent with a perfect adherence record on the implicit-constraint audit task in…
Claude — best write-ups
5.0 / 5clarity
Its final reports read like an analyst's work product: structured what/who/when…
Cursor — cleanest commands
95%accuracy
Across all 180 runs, 5.2% of Cursor's commands errored (40 of its 60 runs were…

Per-metric comparison

Success

cursor
98%
claude
96%
codex
94%

Adherence

cursor
94%
codex
93%
claude
91%

Cmd accuracy

cursor
95%
claude
93%
codex
90%

Tokens

cursor
183,291
codex
279,038
claude
303,862

Agent time

cursor
59s
codex
110s
claude
121s

Cost

cursor
~$0.204
codex
~$0.463
claude
$0.588

Tasks

Overall takeaways

Methodology & grading

Three coding agents — Claude Code (claude-opus-4-8), Codex CLI (gpt-5.6-sol), and Cursor CLI (Composer 2.5 Fast) — ran ten realistic AWS operations tasks against live sandbox accounts, including incident fixes, security remediations, cost cleanups, audits, and forensics.

Test conditions

Each task was run under two conditions to learn when agents benefit from a plan and when they are better left to determine their own approach:

  • Skill: The agent received a detailed, task-specific plan describing how to approach the work.
  • No skill: The agent received only the goal in the prompt and decided how to complete the work on its own.

Repetitions

Each agent completed three repetitions of every task under both conditions, producing 180 graded runs in total.

Grading

Runs were graded in three ways:

  • A deterministic checker inspected live cloud state after each run.
  • An LLM judge graded the agent's process and final report against author-defined criteria.
  • Human review examined every disputed grade.

Human review and overrides

When human review changed a machine grade, the run is labeled Manual override and includes the reason. Several of those corrections became findings in their own right; see the grading-integrity note at the end.