โ— LIVE
OpenAI releases GPT-5 APIIndia AI startup raises $120MBitcoin ETF hits record inflowsMeta Llama 4 benchmarks leakedOpenAI releases GPT-5 APIIndia AI startup raises $120MBitcoin ETF hits record inflowsMeta Llama 4 benchmarks leaked
๐Ÿ“… Sun, 22 Mar, 2026โœˆ๏ธ Telegram
AiFeed24

AI & Tech News

๐Ÿ”
โœˆ๏ธ Follow
๐Ÿ Home๐Ÿค–AI๐Ÿ’ปTech๐Ÿš€Startupsโ‚ฟCrypto๐Ÿ”’Security๐Ÿ‡ฎ๐Ÿ‡ณIndiaโ˜๏ธCloud๐Ÿ”ฅDeals
โœˆ๏ธ News Channel๐Ÿ›’ Deals Channel
Home/Cloud & DevOps/How I Run a Team of AI Coding Agents in Parallel
โ˜๏ธCloud & DevOps

How I Run a Team of AI Coding Agents in Parallel

Running one AI coding agent is productive. Running five in parallel is chaos. I've been using Claude Code daily for months. It's great โ€” until you realize there are four other tasks sitting idle while you wait for one agent to finish refactoring a module. So you open more terminal tabs, spin up more

โšกQuick SummaryAI generating...
B

Batty

๐Ÿ“… Mar 22, 2026ยทโฑ 8 min readยทDev.to โ†—
โœˆ๏ธ Telegram๐• TweetWhatsApp
๐Ÿ“ก

Original Source

Dev.to

https://dev.to/battyterm/how-i-run-a-team-of-ai-coding-agents-in-parallel-1bmn
Read Full โ†—

Running one AI coding agent is productive. Running five in parallel is chaos.

I've been using Claude Code daily for months. It's great โ€” until you realize there are four other tasks sitting idle while you wait for one agent to finish refactoring a module. So you open more terminal tabs, spin up more sessions, and now you're the bottleneck. You're context-switching between agents, resolving merge conflicts they created, and manually checking if anything still compiles.

I spent a week in this mode before I decided there had to be a better way.

The Multi-Agent Problem

Here's what goes wrong when you naively run multiple AI coding agents on the same repo:

They stomp on each other's files. Agent A edits src/auth.rs while Agent B is also editing src/auth.rs. Someone loses.

Nobody checks the tests. An agent says "Done!" but the test suite is failing. You don't find out until three more tasks are stacked on top of the broken one.

You become the dispatcher. Which agent is working on what? Is anyone idle? Did that task actually get assigned? You're doing more coordination than coding.

There's no shared context. Agent A doesn't know Agent B just changed the API interface it depends on. Chaos.

Sound familiar? If you're using Claude Code, Codex, or Aider, and you've ever wanted to run more than one at a time โ€” this is the wall you hit.

How I Solved It

I built Batty โ€” a terminal-native supervisor that turns multiple AI coding agents into a coordinated team. No web UI. No servers. Just your terminal and tmux.

The core idea: instead of a flat pool of agents, you define a hierarchy. An architect agent plans the work. A manager breaks it into tasks. Engineers execute in isolated environments. A kanban board tracks everything. Tests gate completion.

Here's the minimal setup โ€” an architect and three engineers:

# .batty/team_config/team.yaml
name: my-project
board:
  rotation_threshold: 20
standup:
  interval_secs: 600
  output_lines: 40
roles:
  - name: architect
    role_type: architect
    agent: claude
    instances: 1
    prompt: architect.md
    talks_to: [manager]

  - name: manager
    role_type: manager
    agent: claude
    instances: 1
    prompt: manager.md
    talks_to: [architect, engineer]

  - name: engineer
    role_type: engineer
    agent: claude
    instances: 3
    prompt: engineer.md
    talks_to: [manager]
    use_worktrees: true

Three lines that matter:

  1. talks_to โ€” Agents can only communicate with their defined contacts. No free-for-all. The architect talks to the manager, the manager talks to engineers. This prevents the message chaos that kills multi-agent setups.

  2. instances: 3 โ€” Three engineer agents, each in its own tmux pane. Batty names them eng-1-1, eng-1-2, eng-1-3 and manages them independently.

  3. use_worktrees: true โ€” Each engineer works in an isolated git worktree. Their own branch, their own working directory. No merge conflicts during active work.

The Workflow

cargo install kanban-md --locked && cargo install batty-cli
cd my-project
batty init --template simple
batty start --attach

This launches a tmux session. Each agent gets its own pane โ€” you can watch them all work simultaneously, or detach and come back later.

Then you send a task:

batty send architect "Build a REST API with JWT auth and user registration"

Here's what happens:

  1. The architect analyzes the request and breaks it into subtasks
  2. Tasks land on the kanban board (a Markdown file โ€” yes, you can cat it)
  3. The manager dispatches tasks to available engineers
  4. Each engineer picks up a task, creates a branch in its worktree, and starts coding
  5. When an engineer says it's done, Batty runs the test suite
  6. If tests pass, the work is ready to merge. If not, the task goes back to the engineer.

The whole thing is file-based. YAML config, Markdown kanban, Maildir-style inboxes, JSONL event logs. You can git diff your team's entire state.

What I Learned Running This Setup

After a few weeks of daily use, here's what surprised me:

Five parallel agents is the sweet spot

For most repos, 3-5 engineers is ideal. Beyond that, you start hitting genuine merge complexity even with worktree isolation. The agents aren't the bottleneck โ€” the codebase's ability to absorb parallel changes is.

The architect matters more than the engineers

Task decomposition quality is everything. A good architect agent that breaks "Build auth system" into well-scoped, independent subtasks will outperform six engineers working on poorly defined work. I spent more time refining my architect.md prompt than any other part of the setup.

Test gating is non-negotiable

Before Batty, I'd have agents "complete" tasks that broke everything downstream. Now, a task isn't done until tests pass. Period. This single constraint eliminated most of the chaos.

It sounds obvious. But when you're watching five agents work in parallel and one of them says "Done!", the temptation to just accept it and move on is strong. Don't.

You still need to supervise

Batty is not "fire and forget." It's closer to managing a junior dev team than doing the work yourself. You review architecture decisions, redirect when an agent goes off-track, and unblock when someone gets stuck. But you're supervising five workstreams instead of doing one โ€” that's the leverage.

The tmux-native approach just works

I tried web-based dashboards. I tried custom UIs. Nothing beat having the agents in tmux panes where I already work. I can split, resize, scroll back through an agent's history, or detach the whole session and come back from my phone via SSH.

A Real Example

I used Batty to build chess_test โ€” a chess engine built entirely by a team of AI agents. The challenge: build an engine that can beat Stockfish at 1200 ELO at least 50% of the time. No external libraries. No internet lookups.

The team had an architect planning the engine architecture, a manager coordinating the work, and multiple engineers implementing different components in parallel โ€” move generation, evaluation, search algorithms. Each working in their own worktree, each gated on tests.

It's the kind of project that would take one agent days of sequential work. With a coordinated team, the parallel execution compressed the timeline dramatically.

Getting Started

Batty works with the AI coding agents you already use:

  • Claude Code โ€” First-class support, built-in templates
  • Codex โ€” Works as an engineer agent
  • Aider โ€” Works as an engineer agent
  • Custom โ€” Any CLI tool that accepts stdin
# Install
cargo install kanban-md --locked
cargo install batty-cli

# Initialize in your project
cd your-project
batty init --template pair  # start small: 1 architect + 1 engineer

# Launch
batty start --attach

# Send a task
batty send architect "Implement user authentication with JWT"

Eight built-in templates range from solo (one agent, no hierarchy) to large (19 agents with three management layers). Start with pair or simple and scale up as you get comfortable.

What Batty Is Not

I want to be honest about limitations:

  • It's early. Version 0.1.0. The core loop is solid, but the API is still settling.
  • It's not magic. You still need good prompts and good task decomposition. Batty orchestrates โ€” it doesn't think for you.
  • It requires tmux. If you don't use a terminal-based workflow, this isn't your tool.
  • It's not a framework. You can't embed it in your app. It's a CLI supervisor.

If you want a GUI, check out vibe-kanban. If you want a single-agent experience, Claude Code alone is excellent. Batty fills the gap between "one great agent" and "a coordinated team."

Batty is open source, built in Rust, and published on crates.io.

GitHub: github.com/battysh/batty
Demo video: 2-minute walkthrough
Docs: battysh.github.io/batty

If you're already running multiple AI agents and feeling the coordination pain, give it a try. And if you have ideas or feedback โ€” issues and PRs are welcome.

Tags:#cloud#dev.to

Found this useful? Share it!

โœˆ๏ธ Telegram๐• TweetWhatsApp

Read the Full Story

Continue reading on Dev.to

Visit Dev.to โ†—

Related Stories

โ˜๏ธ
โ˜๏ธCloud & DevOps

Why Network as a Service (NaaS) is the Future for Distributed Enterprises

about 2 hours ago

โ˜๏ธ
โ˜๏ธCloud & DevOps

Day 15 โ€“ Building Your First Simple AI Agent

about 2 hours ago

๊ณ ์–‘์ด ๋–จ์–ด์ง€๋Š” ์›๋ฆฌ ๊ณผํ•™์  ์„ค๋ช…๊ณผ ์•ˆ์ „ํ•œ ์ˆ˜์ง๊ณต๊ฐ„ ํ™œ์šฉ๋ฒ•
โ˜๏ธCloud & DevOps

๊ณ ์–‘์ด ๋–จ์–ด์ง€๋Š” ์›๋ฆฌ ๊ณผํ•™์  ์„ค๋ช…๊ณผ ์•ˆ์ „ํ•œ ์ˆ˜์ง๊ณต๊ฐ„ ํ™œ์šฉ๋ฒ•

about 2 hours ago

What I Learned Automating Software Development (After 20 Years of Doing It Manually)
โ˜๏ธCloud & DevOps

What I Learned Automating Software Development (After 20 Years of Doing It Manually)

about 2 hours ago

๐Ÿ“ก Source Details

Dev.to

๐Ÿ“… Mar 22, 2026

๐Ÿ• about 2 hours ago

โฑ 8 min read

๐Ÿ—‚ Cloud & DevOps

Read Original โ†—

Web Hosting

๐ŸŒ Hostinger โ€” 80% Off Hosting

Start your website for โ‚น69/mo. Free domain + SSL included.

Claim Deal โ†’

๐Ÿ“ฌ AiFeed24 Daily

Top 5 AI & tech stories every morning. Join 40,000+ readers.

โœฆ 40,218 subscribers ยท No spam, ever

Cloud Hosting

โ˜๏ธ Vultr โ€” $100 Free Credit

Deploy cloud servers in 25+ locations. From $2.50/mo. No contract.

Claim $100 Credit โ†’
AiFeed24

India's AI-powered tech news hub. Daily coverage of AI, startups, crypto and emerging technology.

โœˆ๏ธ๐Ÿ›’

Topics

Artificial IntelligenceStartups & VCCryptocurrencyCybersecurityCloud & DevOpsIndia Tech

Company

About AiFeed24Write For UsContact

Daily Digest

Top 5 AI stories every morning. 40,000+ readers.

No spam, ever.

ยฉ 2026 AiFeed24 Media.Affiliate Disclosure โ€” We earn commission on qualifying purchases at no extra cost to you.
PrivacyTermsCookies