JaJa: Job Agency Harness news update news update

Open-source developer Kahl-d released jaja, an open-source job-search agency harness for Claude Code that deploys coordinated agent teams to discover roles, evaluate postings, tailor documents, and draft outreach under mandatory human sign-off. According to github.com's report, the software isolates personal candidate records inside a dedicated local directory rather than the public repository. The project packages 30 specialized role agents, 52 command skills, 9 guard hooks, and a 24-tool Model Context Protocol server to automate search logistics while candidate records remain private.

This report details the system architecture, safety controls, and target user base for the jaja job-search framework. Engineered for technical professionals who use Claude Code, the system establishes an auditable local pipeline for multi-stage career searches. Instead of relying on unsupervised application bots, the framework uses deterministic scripts for factual verification while models draft documents and evaluate opportunities under explicit user sign-off.

Job Agency Harness news update: what actually changed

The release introduces an agentic system constructed from 142 Python scripts containing roughly 76,000 lines of standard library, PyYAML, and requests code. Operating as a Claude Code plugin, the codebase keeps candidate data isolated in a private local folder. Installation is handled by a shell script that deploys synthetic demo data and finishes by running scripts/doctor.py, which verifies zero failures on a clean machine. Candidate profiles store verified accomplishments inside facts.yaml and evidence.yaml, ensuring every tailored document maps directly to an authenticated evidence identifier.

System verification relies on 5,192 automated unit tests across 321 test files alongside 70 dedicated Model Context Protocol tests that execute without network access or live personal records. The architecture maintains separation between plugin logic, runtime queues, and candidate data. A built-in leak guard hook halts any git commit that would stage candidate records into the public repository. Markdown role cards define charters, allowed tools, models, budgets, and autonomy tiers from T0 to T4, while synchronization scripts catch configuration drift across Claude Code and experimental Codex adapters.

Job Agency Harness news update: how it works

The runtime architecture organizes 30 role agents across seven operational teams named Find, Judge, Write, Reach out, Track, Prepare, and Run the day. During discovery, agents including job-scout, platform-scout, inbox-harvester, and company-researcher gather postings, inspect platform rules, parse recruiter correspondence, and build cited dossiers. The judgment group uses fit-scorer, role-analyst, legal-visa, and prospect-triage to analyze job postings into ranked technical needs, check candidate qualifications against documented evidence, monitor visa timelines, and score contacts by warmth. Agents execute across eight lanes, accessing 52 skills via slash commands in Claude Code.

Document generation is managed by a ten-agent writing group that drafts application materials while preventing hallucinations. When resume-tailor produces a customized variant, resume-editor verifies claims against pinned evidence records, and base-comparer runs a blind comparison against a hash-pinned base resume, discarding variants that fail to beat the original. Technical reviewers check claim precision, while simulated hiring manager and technical recruiter personas inspect presentation. Outbound messaging halts at an approval queue, ensuring the sender agent only transmits communications that possess an explicit, signed user approval file written to the local runtime folder.

Job Agency Harness news update: why it matters now

The framework addresses growing frustration with indiscriminate job-application bots that flood hiring portals with unreviewed, inaccurate resumes. By using deterministic scripts for counting and gating while restricting models to drafting and evaluation, the system prevents common hallucinations in career documents. Gating checks fail with the exact line quoted from candidate records, enabling users to verify discrepancies or issue direct overrides. Because every factual claim must cite an evidence identifier in local YAML files, candidates avoid fabricated technical qualifications. Sixteen automated background routines ship disabled by default, preventing any unmonitored autonomous execution.

Data ownership and candidate safety differentiate the project from cloud-hosted job-search services. The software enforces pacing caps, operational delays, and automated stops when encountering block pages, pausing all lanes if platform friction arises. Optional external connectors operate in read-only mode by default, requiring separate authentication tokens for outbound tasks. Every action produces immutable run receipts, structured rulings, and append-only event journals inside the local directory. A continuous learning loop evaluates application outcomes across recorded sample sizes, delivering evidence-based proposals to the central hub rather than altering candidate profile records without explicit direction.

Job Agency Harness news update: who is affected

The software is designed for engineers and technical professionals who operate on macOS or Linux systems and use Claude Code for daily workflows. Job seekers managing multi-stage searches across dozens of prospective employers benefit from unified application tracking, contact triage, and systematic interview preparation without storing private career records on third-party cloud servers. Operating the harness requires Python 3.11 or higher, with uv recommended for running the local Model Context Protocol server. Optional dependencies enable integrations with Gmail, Google Calendar, resume typesetting tools, and local PDF rendering utilities for candidates seeking desktop automation.

In contrast, the framework is unsuitable for candidates seeking fully autonomous application bots that dispatch generic resumes without oversight. The architecture intentionally rejects unreviewed submissions, halting every outbound message, email, or application package until an explicit approval file is approved by the user. Setting up the environment requires an afternoon of effort to configure directories, populate factual background records, and organize evidence identifiers. Job seekers wanting simple browser extensions or those unwilling to maintain local YAML files will find the technical prerequisites and strict approval gates incompatible with high-volume, unfocused job hunting.

Job Agency Harness news update: what to watch

Users adopting the harness should monitor schema stability and adapter compatibility as the open-source codebase develops. The author notes that the project remains in an early phase, meaning data layouts, directory structures, and file formats may change across future updates. While Claude Code integration is tested end to end, generated adapters for Codex currently lack comprehensive runtime validation in production environments. Technical candidates and contributors should observe whether upcoming versions preserve backward compatibility across customized agent cards, specialized prompt definitions, and custom skill procedures when updating the core repository files.

Another critical area to monitor is connector compliance and platform safety across third-party communication channels. Integrations for Gmail, Google Calendar, X, and authenticated browser sessions remain optional and disabled by default, placing full responsibility on candidates to respect individual platform terms of service. Users should track the activity of the ops-watchdog agent, which flags silence periods and system health issues to ensure routines remain operational. Over extended searches, candidate results will show whether the hub's outcome learning loop delivers measurable improvements in interview invitation rates and overall job placement efficiency.