Run 1,000 AI Agents in Parallel

Sequential agent execution is a bottleneck. Army Mode fans out tasks to thousands of parallel agents, collects verified results, and returns aggregated output. Turn hours of serial work into minutes.

Why parallel agents matter

A single agent processing 1,000 research queries takes hours. Ten agents take a tenth of the time. Ten thousand agents finish in seconds. The bottleneck in most agent workloads is not intelligence -- it is throughput. Army Mode removes that bottleneck.

Common use cases for massive parallelism:

• Research at scale -- analyze thousands of URLs, papers, or documents simultaneously
• Data processing -- transform, clean, and validate large datasets in parallel
• Testing and verification -- run the same prompt through multiple models or configurations
• Content generation -- produce hundreds of variations, descriptions, or summaries at once
• Competitive analysis -- monitor thousands of competitors, products, or markets concurrently

How Army Mode works

1. Define the task template and the list of inputs (up to 1,000 items)
2. Army Mode dispatches each input to a separate agent executing in parallel
3. Each agent runs the task, calls any Slopshop tools it needs, and returns a result
4. Results are verified against your success criteria (schema validation, content checks)
5. Failed tasks are automatically retried up to 3 times
6. Aggregated results are returned as a single structured response

Built-in verification

Parallel execution is useless if you cannot trust the results. Army Mode includes built-in verification: define a JSON schema for expected output, and each result is validated before inclusion. Results that fail validation are retried automatically. You get a final report showing success rate, failures, and retry counts.

This is critical for production workloads where you need every result to be usable, not just most of them.

Progress tracking and partial results

For long-running Army deployments, you can poll for progress. The API returns completed count, pending count, failed count, and partial results as they come in. Build dashboards, set up alerts, or stream results to downstream systems in real time.

Cost control

Each parallel agent consumes credits for the tools it calls. Memory operations remain free even at scale. You can set a total credit budget for the Army deployment -- if the budget is exhausted, remaining tasks are paused rather than failing silently. This prevents runaway costs on large deployments. See pricing details.

For most workloads, the per-task cost is 1-5 credits, making even a 1,000-task deployment affordable. Compare that to running 1,000 LLM calls through OpenAI -- the tool execution cost is a fraction of the inference cost.