Skip to content

Commit 646d1ce

Browse files
charlypolycharlypoly
committed
feat(examples): add 2 LangChain Deep Agents cookbooks
1 parent 36208f5 commit 646d1ce

22 files changed

Lines changed: 1931 additions & 2 deletions

‎examples/integrations/langchain/README.md‎

Lines changed: 11 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -2,12 +2,21 @@
22

33
This directory contains examples of integrating Langchain with our web automation tools:
44

5-
1. **Browserbase Integration**: A lightweight solution for web scraping and data extraction using our managed browser infrastructure.
5+
1. **Browserbase Integration** (`browserbase/`): A lightweight solution for web scraping and data extraction using our managed browser infrastructure.
66

7-
2. **Stagehand Integration**: Full web automation capabilities using our open-source AI-powered browser automation SDK.
7+
2. **Stagehand Integration** (`stagehand/`): Full web automation capabilities using our open-source AI-powered browser automation SDK.
8+
9+
3. **Deep Agents + Browserbase** (`deepagents-browserbase/`): A Python example combining LangChain Deep Agents with Browserbase Search, Fetch, and Stagehand browser sessions for research workflows.
10+
11+
4. **KYC Onboarding Agent** (`kyc-onboarding/`): A Python example that runs automated KYC due diligence on a company using a five-track fan-out (corporate registry, beneficial ownership, sanctions, litigation, adverse media) and per-owner individual KYC checks. Uses Deep Agents for orchestration and Stagehand for portal navigation.
12+
13+
5. **Patent Landscape Agent** (`patent-landscape-agent/`): A Python example that researches a patent landscape across USPTO, EPO, and WIPO using a three-phase workflow — parallel five-track research (granted patents, prosecution history, ownership, PTAB litigation, inventor network), per-family deep dives, and final memo synthesis. Uses OpenAI as the single model provider for both the Deep Agents orchestrator and Stagehand browser sessions.
814

915
Choose the example that best fits your needs:
1016
- Use Browserbase for simple web scraping and data collection
1117
- Use Stagehand for complex automation workflows with AI-driven interactions
18+
- Use Deep Agents + Browserbase for multi-step research agents with human approval gates
19+
- Use KYC Onboarding for compliance workflows that navigate gated portals and forms
20+
- Use Patent Landscape Agent for multi-source research with persistent workpapers and structured memo output
1221

1322
See the respective directories for detailed implementation guides.

‎examples/integrations/langchain/kyc-onboarding/.env.example‎

Lines changed: 12 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,12 @@
1+
BROWSERBASE_API_KEY=bb_...
2+
3+
# Use direct OpenAI
4+
OPENAI_API_KEY=sk-...
5+
6+
# Or use the Browserbase Model Gateway instead (no separate OpenAI key needed)
7+
# DEEPAGENT_BASE_URL=https://<your-gateway>
8+
9+
# Optional model overrides
10+
# DEEPAGENT_MODEL=gpt-5.4
11+
# STAGEHAND_MODEL=google/gemini-3-flash-preview
12+
# STAGEHAND_AGENT_MODEL=anthropic/claude-sonnet-4-6

‎examples/integrations/langchain/kyc-onboarding/.gitignore‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,4 @@
1+
.env
2+
.venv/
3+
__pycache__/
4+
*.pyc

‎examples/integrations/langchain/kyc-onboarding/README.md‎

Lines changed: 142 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,142 @@
1+
# LangChain Deep Agents + KYC Onboarding (Python)
2+
3+
This example builds a KYC onboarding agent that runs automated due diligence on a company
4+
using LangChain Deep Agents for orchestration, Stagehand for browser interactions, and
5+
Browserbase for headless browser sessions.
6+
7+
The agent navigates real portals — Secretary of State sites, FinCEN BOI, OFAC, PACER, SEC EDGAR,
8+
and news publishers — rather than relying on search or fetch APIs alone. Browserbase Agent Identity
9+
handles anti-bot measures and CAPTCHAs automatically.
10+
11+
## Architecture
12+
13+
**Phase 1 — five-track fan-out:**
14+
15+
| Subagent | Portals |
16+
|---|---|
17+
| `corporate-registry` | Secretary of State portals (Delaware, California, New York, UK Companies House) |
18+
| `beneficial-ownership` | FinCEN BOI, UK PSC register, EU member-state registries |
19+
| `sanctions-pep` | OFAC SDN, EU consolidated list, UK HMT financial sanctions |
20+
| `litigation-regulatory` | PACER, state courts, SEC EDGAR |
21+
| `adverse-media` | Reuters, Bloomberg, FT, WSJ, SEC newsroom, DOJ press releases |
22+
23+
**Phase 2 — individual KYC fan-out:**
24+
25+
After the `beneficial-ownership` subagent returns a list of ultimate beneficial owners (UBOs),
26+
the orchestrator delegates one `individual-kyc` subagent call per person. Each individual check
27+
covers sanctions screening, PEP status, and adverse media.
28+
29+
**Tools (defined in `kyc_tools.py`):**
30+
31+
- `kyc_search`: Browserbase Search for initial discovery before opening portals
32+
- `kyc_fetch`: Browserbase Fetch for static pages (EDGAR filings, public registries)
33+
- `kyc_extract_from_portal`: opens a Browserbase session and extracts structured data from a
34+
rendered portal page using Stagehand `extract`
35+
- `kyc_search_portal`: opens a Browserbase session and runs a multi-step browser task (form fills,
36+
paginated search navigation) using a Stagehand agent — gated behind human approval
37+
38+
Every browser session produces a session replay URL included in the tool response for audit purposes.
39+
40+
## Requirements
41+
42+
- Python 3.11+
43+
- `BROWSERBASE_API_KEY` for Browserbase Search, Fetch, and browser sessions (including Stagehand)
44+
- An OpenAI-compatible model for the Deep Agent orchestrator
45+
46+
The sample accepts `OPENAI_API_KEY` for direct OpenAI access, or `BROWSERBASE_API_KEY` paired with
47+
`DEEPAGENT_BASE_URL` for a Browserbase-backed OpenAI-compatible gateway.
48+
49+
Default models:
50+
- Deep Agent orchestrator: `gpt-4o`
51+
- Stagehand extraction: `google/gemini-3-flash-preview`
52+
- Stagehand interactive agent: `anthropic/claude-sonnet-4-6`
53+
54+
## Install
55+
56+
```bash
57+
cd examples/integrations/langchain/kyc-onboarding
58+
python3 -m venv .venv
59+
source .venv/bin/activate
60+
pip install -r requirements.txt
61+
```
62+
63+
## Environment
64+
65+
Copy `.env.example` to `.env` and fill in your keys:
66+
67+
```bash
68+
cp .env.example .env
69+
```
70+
71+
```ini
72+
BROWSERBASE_API_KEY=bb_...
73+
74+
# Use direct OpenAI
75+
OPENAI_API_KEY=sk-...
76+
77+
# Or use the Browserbase Model Gateway instead (no separate OpenAI key needed)
78+
# DEEPAGENT_BASE_URL=https://<your-gateway>
79+
80+
# Optional model overrides
81+
# DEEPAGENT_MODEL=gpt-5.4
82+
# STAGEHAND_MODEL=google/gemini-3-flash-preview
83+
# STAGEHAND_AGENT_MODEL=anthropic/claude-sonnet-4-6
84+
```
85+
86+
## Run
87+
88+
Use the default demo company:
89+
90+
```bash
91+
python main.py
92+
```
93+
94+
Run on a specific company and jurisdiction:
95+
96+
```bash
97+
python main.py "Stripe, Inc." --jurisdiction Delaware
98+
python main.py "Revolut Ltd" --jurisdiction "United Kingdom"
99+
```
100+
101+
Override the model:
102+
103+
```bash
104+
python main.py "Acme Corp Inc." --jurisdiction Delaware --model gpt-4o
105+
```
106+
107+
## Approval flow
108+
109+
`kyc_search_portal` is gated behind `interrupt_on`. Whenever the agent wants to submit a search
110+
form, navigate paginated results, or interact with a portal, the script pauses and prompts:
111+
112+
```
113+
Pending tool call
114+
Tool: kyc_search_portal
115+
Arguments:
116+
{
117+
"start_url": "https://icis.corp.delaware.gov/ecorp/entitysearch/namesearch.aspx",
118+
"task": "Search for 'Acme Corp Inc.' and extract the entity status, formation date, and registered agent."
119+
}
120+
Allowed decisions: approve, edit, reject
121+
Decision [approve/edit/reject]:
122+
```
123+
124+
- `approve` — run the task as proposed
125+
- `edit` — provide replacement JSON args before running
126+
- `reject` — skip this action
127+
128+
This puts human review at the tool boundary rather than inside ad hoc shell calls.
129+
130+
## Notes
131+
132+
- `kyc_extract_from_portal` maps to Stagehand `extract` — structured, read-only extraction from a rendered page.
133+
- `kyc_search_portal` maps to Stagehand `agent().execute()` — multi-step agentic browsing for portals that require interaction.
134+
- Browserbase Agent Identity manages browser fingerprinting, residential proxy routing, and CAPTCHA solving. No additional configuration is needed beyond `BROWSERBASE_API_KEY`.
135+
- Each `kyc_search_portal` and `kyc_extract_from_portal` response includes a `session_url` field linking to the Browserbase session replay. These replays provide visual evidence for audit workpapers.
136+
- The individual-kyc Phase 2 fan-out is driven by the orchestrator parsing the `beneficial-ownership` subagent report. If the registry is not publicly searchable, the subagent notes this explicitly and Phase 2 proceeds with whatever names were found.
137+
138+
## Suggested prompts
139+
140+
- `python main.py "Anthropic, PBC" --jurisdiction Delaware`
141+
- `python main.py "Klarna Bank AB" --jurisdiction Sweden`
142+
- `python main.py "Acme Corp Inc." --jurisdiction Delaware` (demo company, expect sparse results)

0 commit comments

Comments
 (0)