atq

Batch process tasks using agents in parallel.

The problem

You have a large batch of tasks that need to be processed using an LLM agent (language/reasoning/tool use tasks).

Running these through a single long-lived agent doesn't work:

• Agents usually want to write a script to process batch tasks - rather than reason through each one.
• Context bloat. The agent accumulates results from previous tasks. Gets slower, more expensive, less focused.
• Fragile batching. One failure midway can stall everything. Tracking what succeeded is painful.
• No concurrency. Processing items one at a time when each task is independent is just slow.

The approach

Isolate each with a clean context. A pool of agents process them in parallel.

• Fresh context per task. Each agent only sees the item it's working on. Better focus, lower cost.
• Concurrent by default. Control the pool size with concurrency.
• Stream results. Results stream to stdout as agents complete.

CLI

npm install -g @endyai/atq

cat companies.txt | atq "Find the current CEO of this company. Return their full name." -c 10

Input is piped via stdin (one line per item):

Apple
Google
Meta

Output (stdout, one line per item, in input order):

Tim Cook
Sundar Pichai
Mark Zuckerberg

Progress (stderr):

[1/3]
[2/3]
[3/3]

Flags

| Short | Long | Required | Default | Description | | ----- | ----------------- | -------- | -------------------- | ----------------------------------------------------- | | -f | --prompt-file | — | — | Read prompt from a file | | -c | --concurrency | no | 10 | Max parallel agents | | -p | --provider | no | anthropic | Provider: anthropic or openai | | -m | --model | no | per provider | Model name | | -k | --api-key | no | — | API key (or set ANTHROPIC_API_KEY / OPENAI_API_KEY) | | -r | --retries | no | 3 | Max retries per failed item (exponential backoff) | | -v | --verbose | no | — | Print agent messages to stderr | | -t | --allowed-tools | no | — | Comma-separated list of tools |

Prompt is passed as a positional argument or via --prompt-file.

Providers

Supports Anthropic and OpenAI agents. Set with --provider or it defaults to anthropic.

| Provider | Default model | SDK | | -------- | -------------------- | ---------------------------- | | anthropic | claude-sonnet-4-6 | @anthropic-ai/claude-agent-sdk | | openai | gpt-5.4 | @openai/agents |

Examples

Clean song titles

cat songs.txt | atq "Clean the song title. Remove featured artists, extra tags like (Official Video), remaster notes, etc. Return just the clean song title." -c 10

Input:

Bohemian Rhapsody (Remastered 2011)
Blinding Lights (feat. Doja Cat) [Official Video]
Hotel California - 2013 Remaster

Output:

Bohemian Rhapsody
Blinding Lights
Hotel California

Find CEO with web search

cat companies.txt | atq "Find the current CEO of this company. Return their full name." -c 5 -t WebSearch

Input:

Rivian
Figma
Stripe

Output:

RJ Scaringe
Dylan Field
Patrick Collison

Research & enrich a database

cat ids.txt | atq "You have a sqlite db at mydb.db. For the given id:
1. Look up the company name from the companies table
2. Search the web for their most recent funding round
3. Download their logo and save it to ./logos/<id>.png
4. Update the company row with last_round, amount, and logo_path" -c 5 -t WebSearch

Input:

a1b2c3d4
e5f6g7h8
i9j0k1l2

Output:

Rippling → Series F, $200M, logo saved
Vercel → Series E, $250M, logo saved
Cursor → Series B, $105M, logo saved

SDK

If you need to run atq programmatically from your own code:

npm install @endyai/atq

import { Task } from '@endyai/atq';

const task = new Task({
  prompt: 'Find the current CEO of this company. Return their full name.',
  concurrency: 10,
  items: ['Apple', 'Google', 'Meta'],
});

for await (const { item, output, progress } of task.run()) {
  console.log(`[${progress.completed}/${progress.total}] ${item} → ${output}`);
}

new Task(options)

| Option | Type | Default | Description | | -------------- | ---------- | -------------------- | -------------------------------- | | prompt | string | — | Instructions for the agent | | concurrency | number | 10 | Max parallel agents | | retries | number | 3 | Max retries per failed item | | verbose | boolean | false | Print agent messages to stderr | | provider | string | 'claude' | 'claude' or 'openai' | | items | array | [] | Items to process | | model | string | per provider | Model name | | apiKey | string | — | API key for the provider | | allowedTools | string[] | — | Tools the agent can use |

.run()

Async generator. Each yield: { item, output, progress: { completed, total } }