THE OPEN WORKBENCH / 01
Run it. Break it. Understand what happened.
Route map: evidence leads to draft, then check. A failed check loops through revise or stops. A passing check reaches review, then finish.
Every box is a function. A box becomes an agent only when it can make model-driven decisions or use tools autonomously.
WHY IT EXISTS
THE HANDOFF
Run this node to inspect its actual state.
Excerpt from the runnable Python implementation.
Download full workflow.pyCHECKPOINT / WAITING FOR YOU
Its structure and citation IDs passed. You still need to judge whether the claims are useful and supported.
This completes a practice run. Nothing is sent or published.
Your first run starts here. Try Step once and watch the shared state change.
FROM DIAGRAM TO YOUR OWN CODE
A small program you can read end to end. It uses actual LangGraph, a replaceable writer function, and a SQLite checkpointer. The default writer is deterministic; no API key is needed.
Download Python starterUnzip the starter. Open a terminal in that folder. Use Python 3.12, the version used for verification.
python -m venv .venv # macOS / Linux source .venv/bin/activate # Windows PowerShell instead: # .\.venv\Scripts\Activate.ps1 python -m pip install -r requirements.txt python workflow.py --config config.json --thread first-run
The program prints the actual input, patch, and output for each completed node, then pauses for review.
Close the Python process. From the same directory, use the same database and thread ID:
python workflow.py --thread first-run --decision approve # Or use --decision reject
SQLite preserves the checkpoint in workshop.sqlite. Browser checkpoints last only until this page reloads; the Python version demonstrates durable resume.
Use a new thread ID for each new experiment. A thread ID is a lookup key, not authentication.
config.json.sources.json. Keep IDs unique.retrieve() to search your own small dataset.validate_draft() independent of the writer.python -m unittest discover -s tests -v
Try a retrieval limit of 1 with a minimum of 4 citations. A writer cannot repair evidence it never received. The next improvement is a route back to retrieval.
build_graph(..., writer=your_writer) is the model boundary. Your function receives the selected evidence and validation feedback, then returns a list of {text, source_id} objects.
Assemble instructions separately from untrusted source text. Require structured output, validate it, cap input and output, add a timeout, and count every attempted model call. The model does not get to approve itself.
Live Osiris subscription inference is not enabled in this lab. It needs the approved hosted connection, this feature’s own server-side contract, consent and usage controls. Signing in alone is not permission to generate.
Open to load workflow.py.
THE PARTS, AND WHY THEY MATTER
The graph decides what runs next. The model supplies judgment where you deliberately allow it. Your application owns state, permissions, evidence, and stopping.
Define the input, shared state, and expected output before writing prompts. A Python TypedDict documents shape; it does not validate model output at runtime. Validators do that work.
Store evidence and decisions as structured fields. Pass each node only what it needs. A checkpoint is execution memory, not a user profile or a vector database.
A node is a function. A tool does a bounded job such as retrieval or a calculation. An agent uses a model to decide actions in a loop. Giving five functions job titles does not create five intelligent coworkers.
This lab is a controlled workflow with a practice writer. Start here; add autonomous tool selection only where it earns its complexity.
Conditional edges turn evidence into control flow. No evidence stops the run. Invalid output gets a limited repair attempt. Valid output reaches a person.
A repair changes the content using feedback. A transport retry repeats a failed request, usually with backoff. This lab teaches repair; it makes no network requests to a provider.
The browser pauses at LangGraph node boundaries and resumes with updateState(). The Python version uses interrupt() and Command(resume=...) with durable SQLite storage.
An interrupted Python node restarts on resume. Keep external effects after approval and use an idempotency key so a retry cannot publish or charge twice.
Observe inputs, outputs, tool results, routing, errors, latency, and usage. Those are execution records, not a model’s hidden reasoning.
Keep test cases for good evidence, unknown citations, empty retrieval, failed repair, approval, rejection, and recovery after a restart. A second model’s opinion is another signal, not ground truth.
Keep the interface on the site. Run live graphs behind an authenticated service with per-user checkpoint ownership, bounded concurrency, timeouts, cancellation, and usage limits. Never execute a visitor’s arbitrary Python on the public server.
Start with this fixed graph and editable configuration. Python needs a Python-capable host; a JavaScript graph needs a compatible runtime. LangGraph does not require DigitalOcean or its own managed hosting.
| Layer | This release | Live service needs |
|---|---|---|
| Interface | Graph, inspector, experiments, trace export | Authenticated run controls and streamed events |
| Orchestration | Browser LangGraph + matching Python starter | Server-enforced rules and per-user isolation |
| Intelligence | Deterministic practice writer | Approved visitor-funded Osiris connection |
| Persistence | Page memory; SQLite in Python | Durable database with ownership checks |
| Actions | Draft approval/rejection only | Allowlisted tools and effect-specific approval |
As checked October 8, 2026, Osiris’s hosted ChatGPT plan adapter remains in draft PR #191. This workshop neither activates it nor routes usage to an operator-funded API.
Plain Python first for a short linear sequence. LangGraph here because branching, shared state, repair, and pause/resume are the lesson. A drag-and-drop editor can come later; understanding the runtime comes first.
Next build challenge: add a retrieval-repair branch when the draft needs more evidence, with a separate retrieval budget. After that, try two specialists in parallel and define how their state updates merge.