Add LangChain Deep Agents + MongoDB Atlas VFS tutorial

partners/langchain/langchain-deepagents/.env.example

@@ -0,0 +1,24 @@
+# MongoDB Atlas — M0 works for dev; M10+ for Search/Vector Search in production
+MONGODB_URI=mongodb+srv://<user>:<pass>@cluster.mongodb.net/?appName=devrel-tutorial-deepagents-langchain-vfs
+
+# AWS — S3 bucket and region (region governs S3, SQS, AND Bedrock together)
+S3_BUCKET_NAME=my-deepagents-vfs-bucket
+AWS_REGION=us-east-1
+AWS_ACCESS_KEY_ID=
+AWS_SECRET_ACCESS_KEY=
+
+# Embedding provider: "bedrock" (default) or "openai"
+# Bedrock uses Titan Text Embeddings v2 — must be enabled in your region.
+# OpenAI drops the Bedrock enablement step at the cost of a second API key.
+EMBEDDING_PROVIDER=openai
+
+# Required when EMBEDDING_PROVIDER=openai
+OPENAI_API_KEY=sk-...
+
+# LLM for agents (OpenAI)
+# Used by the coordinator and sub-agents for reasoning.
+# OPENAI_API_KEY above covers this if using OpenAI for both.
+
+# regulations.gov API key — only needed for tools/fetch_docket.py (one-time corpus pull)
+# NOT needed to run the demo; the corpus is vendored in corpus/.
+REGULATIONS_GOV_API_KEY=

partners/langchain/langchain-deepagents/.gitignore

@@ -0,0 +1,30 @@
+# Python
+__pycache__/
+*.pyc
+*.egg-info/
+dist/
+build/
+.eggs/
+
+# Environment
+.env
+.venv/
+venv/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+
+# Vendored source (reference only, not shipped)
+_vendored_vfs/
+
+# Claude Code
+CLAUDE.md
+
+# Internal feedback (not for publication)
+INTEGRATION_FEEDBACK.md
+
+# OS
+.DS_Store
+Thumbs.db

partners/langchain/langchain-deepagents/README.md

@@ -0,0 +1,126 @@
+# MongoDB Atlas VFS for LangChain Deep Agents
+
+**Building a Multi-Agent Pipeline Where Nothing Gets Lost**
+
+A multi-agent research pipeline that reads a real federal rulemaking docket
+(DOE air-cleaner efficiency standards), finds cross-document discrepancies no
+single file contains, and survives being killed mid-run.
+
+Demonstrates [langchain-mongodb-deepagents-vfs](https://github.com/langchain-ai/langchain-mongodb) —
+a MongoDB Atlas-backed virtual filesystem adapter for
+[LangChain Deep Agents](https://github.com/langchain-ai/deepagents).
+
+## What this demo shows
+
+1. **Discovery across formats** — hybrid search (`$rankFusion`) finds the same
+   concept across PDFs, XLSX spreadsheets, and DOCX uploads where `ripgrep` and
+   literal-substring search cannot.
+
+2. **Multi-agent coordination via shared workspace** — four sub-agents search a
+   read-only corpus and write findings to a durable workspace. A writer reads
+   those findings by exact path and produces a final memo.
+
+3. **The kill test** — `SIGKILL` the pipeline after 2 of 4 stages. Resume with
+   the same `run_id`. It picks up where it left off because the workspace survived.
+
+## Prerequisites
+
+**Three accounts required:**
+
+| Service | What for | Tier |
+|---------|----------|------|
+| MongoDB Atlas | Chunk storage + hybrid search | M0 (dev) / M10+ (production) |
+| AWS | S3 bucket + optional Bedrock embeddings | Free tier works for S3 |
+| OpenAI | LLM for agents + embeddings (with `EMBEDDING_PROVIDER=openai`) | Pay-as-you-go |
+
+> **Region gotcha:** `AWS_REGION` governs S3, SQS, **and** Bedrock together —
+> they cannot be split. If using Bedrock embeddings, Titan Text Embeddings v2
+> must be enabled in that region or you'll get `NoRegionError`.
+
+## Setup
+
+```bash
+# Clone and install
+cd partners/langchain/langchain-deepagents
+pip install -e ".[openai,dev]"
+
+# Configure
+cp .env.example .env
+# Edit .env with your credentials
+
+# Seed the corpus into S3 + MongoDB
+python scripts/00_seed_corpus.py
+```
+
+## Running the demo
+
+### Beat 1 — Discovery across formats
+
+```bash
+# Control: ripgrep over local files
+bash scripts/01_control_grep.sh
+
+# Control: StoreBackend + MongoDBStore (literal substring only)
+python scripts/01b_control_storebackend.py
+
+# MongoFilesystemBackend — hybrid $rankFusion search
+python scripts/02_discovery.py
+```
+
+### Beat 2 — Multi-agent pipeline
+
+```bash
+python scripts/03_pipeline.py --run-id aircleaners-001
+```
+
+### Beat 3 — The kill test
+
+```bash
+# Kill after 2 of 4 stages
+python scripts/03_pipeline.py --run-id aircleaners-002 --kill-after 2
+
+# Resume — reads the manifest, skips completed stages
+python scripts/04_resume.py --run-id aircleaners-002
+```
+
+## Architecture
+
+```
+coordinator
+├── writes  workspace/<run_id>/plan.md
+├── writes  workspace/<run_id>/manifest.json       ← run receipt
+├── task → proposal-reader  → findings/proposal.md
+├── task → adoption-reader  → findings/adopted.md
+├── task → numbers-reader   → findings/numbers.md
+└── task → writer           → reads three by path → memo.md
+```
+
+**Two planes, two guarantees:**
+
+| Plane | Path | Operations | Guarantee |
+|-------|------|-----------|-----------|
+| Corpus (discovery) | `corpus/` | `grep`, `glob`, `ls` | Eventually consistent |
+| Workspace (coordination) | `workspace/<run_id>/` | `write`, `read`, `edit` | Read-after-write |
+
+Agents **discover** through the corpus and **coordinate** through the workspace.
+
+## The corpus
+
+DOE docket EERE-2021-BT-STD-0035: Energy Conservation Standards for Air Cleaners.
+Public domain. Vendored in `corpus/` — no API key needed to run the demo.
+
+**The question:** *Did DOE adopt what the Joint Stakeholders proposed, and do the
+energy-savings numbers agree?*
+
+## Cost
+
+Approximate cost per full pipeline run (4 stages, gpt-4o):
+
+| Run type | Tokens | Cost |
+|----------|--------|------|
+| Cold | ~15K–25K | ~$0.05–0.15 |
+| Resumed (after kill at stage 2) | ~8K–12K | ~$0.03–0.07 |
+
+## License
+
+Apache-2.0. Corpus documents are US federal government publications (public domain).

partners/langchain/langchain-deepagents/corpus/README.md

@@ -0,0 +1,50 @@
+# Corpus: DOE Docket EERE-2021-BT-STD-0035
+
+Energy Conservation Standards for Air Cleaners.
+All files are US federal government publications (public domain).
+
+## How to populate
+
+Run `tools/fetch_docket.py` with a regulations.gov API key:
+
+```bash
+export REGULATIONS_GOV_API_KEY=your-key
+python tools/fetch_docket.py
+```
+
+This downloads the full docket into the directory structure below.
+
+## Expected structure
+
+```
+corpus/
+├── rules/
+│   ├── 88FR21752-final-rule.pdf    # Final rule (88 FR 21752)
+│   ├── nopr.pdf                    # Simultaneous NOPR (FR doc 2023-06498)
+│   └── confirmation.pdf            # Confirmation of dates (FR doc 2023-18860)
+├── analysis/
+│   ├── tsd.pdf                     # Technical Support Document
+│   ├── lcc.xlsx                    # Life-Cycle Cost Analysis
+│   ├── nia.xlsx                    # National Impact Analysis
+│   ├── grim-joint.xlsx             # GRIM — Joint Proposal version
+│   └── grim-dfr.xlsx               # GRIM — Direct Final Rule version
+└── comments/
+    ├── 0003-trane.pdf
+    ├── 0005-miaq.pdf
+    ├── 0006-electrolux.pdf
+    ├── 0007-lennox.pdf
+    ├── 0008-joint-commenters.pdf
+    ├── 0009-ca-ious.pdf
+    ├── 0010-blueair.pdf
+    ├── 0011-molekule.pdf
+    ├── 0012-daikin.pdf
+    ├── 0013-neea.pdf
+    ├── 0014-synexis.pdf
+    ├── 0015-ahri.pdf
+    └── 0016-joint-stakeholders.pdf
+```
+
+## Source
+
+- Federal Register: https://www.federalregister.gov/documents/2023/04/11/2023-06499
+- regulations.gov: https://www.regulations.gov/docket/EERE-2021-BT-STD-0035