The problem
Multi-agent demos optimize for the screenshot: agents talk, something plausible comes out, nobody checks it. This system was built to run unattended every day, which forces the opposite priorities — outputs gated on measured accuracy, and LLM spend driven toward zero wherever deterministic code can do the job.
Architecture
Six LangGraph patterns cover the coordination space: hierarchical (supervisor routes to workers), peer debate (agents cross-critique each round), dynamic swarm (task queue with runtime re-analysis), enhanced swarm (swarm plus GRPO scoring, persona evolution, and RLM synthesis), plan-and-execute, and map-reduce. Every pattern ends at the same dual convergence gate: quality score ≥ 8.0/10 AND factual accuracy ≥ 9.5/10, with claim-level verification against research notes, web search, and cached facts.
Fact-checking is deliberately not circular. A summary verified only against its own abstract scores 0.5, not 1.0 — external verification via Semantic Scholar, credibility-weighted web search, and GitHub repo health is what earns a high score, and each verdict ships a human-readable explanation.
The patterns power JobPulse, 15+ agents running daily on cron: Gmail triage, calendar, GitHub activity, arXiv digests, Notion tasks, budget tracking, and a job-application autopilot. The autopilot's pre-screen models a recruiter's 6–30 second scan as four deterministic gates — title relevance, kill signals, must-have coverage, and a 0–100 competitiveness score — over roughly 250 raw jobs a day at zero LLM cost. Per the repo's figures, that cut LLM calls by 96% ($5.63 to $0.23 per month), with a 582-entry skill taxonomy handling extraction and an LLM fallback touching only the ~15% of postings it can't parse.
What's measured
The repo header reports ~146,500 LOC across 684 Python files and 3,458 tests, auto-updated by a stats script rather than typed by hand. Dashboards expose agent success rates, GRPO scores, persona drift, and per-agent cost; retrieval quality is tracked with MRR, NDCG@k, and recall@k.
Honest limitations
This is a single-user production system — it runs one person's life, and the cost and volume figures are that user's telemetry, self-reported by the repo's own scripts rather than independently benchmarked. The accuracy gate measures the fact-checker's scoring pipeline, not ground truth: a claim the web gets wrong can still pass.