# How Batch Processing with Parallel AI Agents Works in Career‑Ops > Discover how Career-Ops uses batch processing and parallel AI agents to efficiently evaluate job postings. Learn about atomic reservations, isolated Node.js workers, and result aggregation. - Repository: [Santiago Fernández de Valderrama/career-ops](https://github.com/santifer/career-ops) - Tags: deep-dive - Published: 2026-08-19 --- **Career‑Ops evaluates multiple job postings simultaneously by atomically reserving unique report numbers, spawning independent Node.js workers that each run the full `auto-pipeline` in isolation, and aggregating the results back into the central tracker after all parallel processes complete.** The `santifer/career-ops` repository implements an embarrassingly parallel architecture for high‑throughput job evaluation. By eliminating shared mutable state and reserving identifiers upfront, the system safely scales batch processing with parallel AI agents across many job descriptions without race conditions or duplicate reports. ## Atomic Report Number Reservation Before any workers start, the system must guarantee that every parallel agent writes to a distinct report ID. The script `reserve-report-num.mjs` handles this by atomically claiming a contiguous block of report numbers—such as `042-049`—and persisting the reservation to a lock file. This reservation step is the integrity layer of the pipeline. Because each worker receives a pre‑allocated range, simultaneous processes can never claim the same identifier. The lock file ensures that even under heavy concurrency, the allocated block remains exclusive to the current batch job. ## Parallel Worker Execution with batch-tailor.mjs Once the report range is reserved, `batch-tailor.mjs` distributes the IDs to independent workers. Each process is launched via a simple Node.js command and operates on its own reserved slice of the range. ```bash node batch-tailor.mjs --start 042 --count 8 ``` In this example, the script spawns workers beginning at report `042` for a total of eight parallel agents. Each worker invokes the standard evaluation mode, referred to in the codebase as `auto-pipeline`, which parses the job description, scores the fit, and generates a PDF. Because every agent writes to its own files—such as `reports/042-*.md` and `output/042-*.pdf`—there is no shared mutable state between processes. ## Result Aggregation and Tracker Consolidation After all parallel agents finish, the pipeline enters a single‑threaded aggregation phase. The script `merge-tracker.mjs` collects the generated TSV additions produced by each worker and integrates them into the canonical [`data/applications.md`](https://github.com/santifer/career-ops/blob/main/data/applications.md) tracker. This step performs consistency checks including deduplication and status validation, ensuring the tracker remains the single source of truth. Token usage across all workers is summed by `batch/aggregate-tokens.mjs`, which produces a concise usage report. This accounting script runs once the parallel jobs are complete, giving an accurate total of API consumption for the entire batch without needing cross‑process communication during execution. ## Summary - **`reserve-report-num.mjs`** atomically reserves a contiguous range of report numbers to prevent race conditions before workers start. - **`batch-tailor.mjs`** launches independent Node.js workers via command‑line flags like `--start` and `--count`, keeping execution embarrassingly parallel. - Each agent runs the full **`auto-pipeline`** in isolation, writing to unique file paths such as `reports/042-*.md` and `output/042-*.pdf`. - **`merge-tracker.mjs`** consolidates worker outputs into [`data/applications.md`](https://github.com/santifer/career-ops/blob/main/data/applications.md) with deduplication and validation. - **`batch/aggregate-tokens.mjs`** tallies total token usage after all workers finish, providing a unified cost report. ## Frequently Asked Questions ### How does Career‑Ops prevent duplicate report numbers during parallel execution? The system calls `reserve-report-num.mjs` before spawning any agents. This script atomically claims a contiguous block of report numbers and writes the reservation to a lock file, guaranteeing that each parallel worker receives a unique identifier and eliminating race conditions. ### What command launches the parallel AI agents in a batch job? You start the workers by running `node batch-tailor.mjs --start --count `. For example, `node batch-tailor.mjs --start 042 --count 8` distributes reserved IDs to eight independent processes that each execute the full evaluation pipeline. ### How are results merged after all parallel agents finish? The script `merge-tracker.mjs` gathers the TSV additions generated by each worker and merges them into [`data/applications.md`](https://github.com/santifer/career-ops/blob/main/data/applications.md). It performs deduplication and status validation to maintain a consistent, canonical tracker. ### Where is token usage aggregated across all batch workers? Token consumption is totaled by `batch/aggregate-tokens.mjs` after all parallel jobs complete. This produces a single usage report without requiring inter‑process communication during the parallel execution phase.