🤖 AutoDebug

A 5-agent autonomous bug detection and reporting pipeline, built natively on Band.

---

What is AutoDebug?

AutoDebug is a 5-agent system that detects bugs in a GitHub repository, analyses the root cause, generates a fix, confirms the bug through testing, and delivers a clear, actionable fix report by email — with zero auto-merging. A human always reviews and applies the fix.

The 5 agents are real, independent processes connected to Band. They coordinate entirely through Band chat rooms using @mentions — Band is not a wrapper here, it's the actual communication layer the agents depend on to hand off work.

---

Architecture

@AutoDebug-Detector detect

↓ (Band room, @mention)

DetectorAgent

clones repo finds failed CI run, or runs files directly to catch the error

↓ @AutoDebug-Analyser (Band)

AnalyserAgent AI root cause analysis, severity rating

↓ @AutoDebug-Fixer (Band)

FixerAgent generates a suggested code fix

↓ @AutoDebug-Tester (Band)

TesterAgent runs pytest, confirms the bug is real

↓ @AutoDebug-Resolver (Band)

ResolverAgent writes a human-readable fix report triggers email + DevOps AI Notifier

Each agent is a standalone Python process built with Band's SimpleAdapter. They run continuously (agent.run_forever()), listening in a shared Band chat room, and hand off work to the next agent by sending an @mention with the accumulated context as JSON.

---

The 5 Agents

Agent	Role	Hands off to
Detector	Finds the bug — from CI logs or by running files directly	Analyser
Analyser	AI-powered root cause analysis and severity rating	Fixer
Fixer	Generates a suggested code fix	Tester
Tester	Runs pytest, confirms the bug is real	Resolver
Resolver	Writes the fix report, sends email, triggers DevOps Notifier	—

---

Key Features

• Real Band-native agents — 5 separate registered agents, coordinating through @mentions in a shared room
• Human-in-the-loop by design — never auto-merges. Every fix is a report, not a commit
• Email delivery — final fix report lands directly in your inbox
• DevOps AI Notifier integration — connects to a second existing project for summarised notifications
• CLI management — autodebug start/stop/status/doctor manages all 5 agents without 5 manual terminals
• Local fallback pipeline — agents/ + core/pipeline.py provide the same 5-stage logic outside Band, useful for offline testing

---

Tech Stack

• Band SDK (band-sdk) — agent registration, rooms, @mention routing
• NVIDIA API (MiniMax M2.7) — reasoning for Analyser, Fixer, Resolver
• Python 3.11+ — all agent logic
• GitHub API — CI log retrieval, repo cloning
• Gmail SMTP — final report delivery
• pytest — bug confirmation

---

Setup

1. Clone and install

git clone https://github.com/Ekaanksh-dev/AutoDebug.git
cd AutoDebug
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install band-sdk

2. Register 5 agents on Band

Go to app.band.ai/agents and create 5 External Agents: AutoDebug-Detector, AutoDebug-Analyser, AutoDebug-Fixer, AutoDebug-Tester, AutoDebug-Resolver.

Copy each Agent UUID + API Key into agent_config.yaml:

detector:
  agent_id: "..."
  api_key: "..."
analyser:
  agent_id: "..."
  api_key: "..."
fixer:
  agent_id: "..."
  api_key: "..."
tester:
  agent_id: "..."
  api_key: "..."
resolver:
  agent_id: "..."
  api_key: "..."

3. Configure .env

cp .env.example .env
nano .env  # fill in your API keys

4. Start all 5 agents

autodebug start
autodebug status

5. Create a Band chat room

In Band, create a chat room and add all 5 agents as participants.

6. Trigger detection

@AutoDebug-Detector detect Ekaanksh-dev/test-buggy-repo 27ac9df0... main

Watch the room — each agent responds and hands off to the next automatically.

---

CLI Reference

Command	Description
autodebug start	Launches all 5 agents as background processes
autodebug stop	Stops all 5 agents
autodebug status	Shows which agents are running
autodebug doctor	Validates config files and agent scripts
autodebug detect <repo> <commit> <branch>	Prints the exact Band mention command to trigger detection

---

Local Fallback Mode (no Band)

The original sequential pipeline still works standalone, useful for local testing without Band:

python main.py analyse --repo owner/repo --commit <sha> --branch main
python main.py test-email

---

Hackathon

Built for the Band of Agents Hackathon (June 12–19, 2026) Track: Multi-Agent Software Development Platform: lablab.ai

---

License

Copyright 2026 Ekaanksh (@Ekaanksh-dev)

Licensed under the Apache License 2.0

---

Built by Ekaanksh | Powered by Band + NVIDIA API