Cyberclaw — Case Study

Overnight Engineering
Guaranteed Correctness

  • Self-orchestrated debugging & repair
  • AI agent reasoning and code fixing
  • Continuous monitoring & validation
  • Full automation from failure to success

Scroll to explore ↓

Step 0 — User Input
System Architecture
Step 0
User Input

"Please fix the deployment so it passes all of the e2e user story tests"

SKILLS ACTIVATE

Agent knows to SSH into relevant dev VM, and reads the relevant SoP restriction that editing e2e user story tests is strictly prohibited

[Internal Thought]: I must prioritize the e2e test failure. Before modifying any code, I will verify the environment constraints to ensure no forbidden files are edited...
Step 1
Orchestration

Agent checking logs of prior engineering run

SKILLS ACTIVATE

Agent knows to commit current repo state if positive progress made — else, it leaves the repo state uncommitted, just 'resetting the deployed infra state for a fresh run' and then spawning another coding agent through automate_loop.sh

[Log]: Analyzing session_logs.jsonl... Previous attempt failed at step 12/13. No commit created to avoid pollution. Resetting infra state...
Step 2
Diagnosis & Fix

Automate_loop.sh executing its 3-stage process

SKILLS ACTIVATE

Calling automate_loop.sh with relevant paths + a succinct summary of the prior agent's goal

$ ./automate_loop.sh --goal "Fix e2e step 13" --context "Root cause: timeout in auth-proxy" [Plan]: 1. Update proxy timeout to 30s | 2. Verify connectivity | 3. Run test
Step 3
Validation

System validates progress through pipeline using solution.sh script

SKILLS ACTIVATE

Orchestrator behavior: if prior coding CLI succeeded, runs solution.sh to determine current progress and saves result in error_output location; if prior coding CLI failed, skips validation as spawned agent will handle it

[Validator]: Executing solution.sh... [Result]: Step 13/13 PASSED — Final state validated.
Output
Success

"Successfully completed 13/13 exploit chain steps after 6+ hours of autonomous operation"

SKILLS ACTIVATE

Agent delivers complete success without human intervention

Total Duration: 6h 24m Iterations: 11 loops Final Status: ALL TESTS PASSED
13/13 Tests Passed
6h+ Autonomous
11× Iterations
Step 0
User Input

"Please fix the deployment so it passes all of the e2e user story tests"

SKILLS ACTIVATE

Agent knows to SSH into relevant dev VM, and reads the relevant SoP restriction that editing e2e user story tests is strictly prohibited

[Internal Thought]: I must prioritize the e2e test failure. Before modifying any code, I will verify the environment constraints to ensure no forbidden files are edited...