Agent Feedback Loop

The core use case for Notebind: your AI agent writes a document, humans review it, and the agent iterates based on feedback.

The workflow

Agent writes markdown
       │
       ▼
  Push to Notebind ──────────▶ Human reviews in browser
       │                              │
       │                    Leaves comments & suggestions
       │                              │
       ▼                              ▼
  Pull feedback ◀──────────── Feedback available via API
       │
       ▼
 Agent processes feedback
       │
       ▼
 Agent updates document
       │
       ▼
  Push updated version ──────▶ Repeat

Implementation

1. Agent creates the document

   import requests
   
   API_KEY = "nb_sk_YOUR_KEY"
   BASE = "https://notebind.com/api"
   headers = {
       "Authorization": f"Bearer {API_KEY}",
       "Content-Type": "application/json"
   }
   
   # Create document
   doc = requests.post(f"{BASE}/documents", headers=headers, json={
       "title": "Q1 Report Draft",
       "content": agent_generated_markdown
   }).json()["data"]
   
   doc_id = doc["id"]

2. Create a share link for reviewers

   share = requests.post(
       f"{BASE}/documents/{doc_id}/share",
       headers=headers,
       json={"permission": "comment"}
   ).json()["data"]
   
   review_url = f"https://notebind.com/share/{share['token']}"
   # Send review_url to human reviewers via email, Slack, etc.

3. Poll for feedback

   import time
   
   def get_unresolved_feedback(doc_id):
       comments = requests.get(
           f"{BASE}/documents/{doc_id}/comments",
           headers=headers
       ).json()["data"]
   
       suggestions = requests.get(
           f"{BASE}/documents/{doc_id}/suggestions",
           headers=headers
       ).json()["data"]
   
       unresolved_comments = [c for c in comments if not c["resolved"]]
       pending_suggestions = [s for s in suggestions if s["status"] == "pending"]
   
       return unresolved_comments, pending_suggestions
   
   # Check periodically
   while True:
       comments, suggestions = get_unresolved_feedback(doc_id)
       if comments or suggestions:
           break
       time.sleep(60)  # Check every minute

4. Process feedback and update

   # Feed comments to your LLM for processing
   feedback_prompt = "Here are the review comments on the document:\n\n"
   for comment in comments:
       feedback_prompt += f"- {comment['body']}"
       if comment.get("anchor_text"):
           feedback_prompt += f" (on: \"{comment['anchor_text']}\")"
       feedback_prompt += "\n"
   
   for suggestion in suggestions:
       feedback_prompt += f"- Replace \"{suggestion['original_text']}\" → \"{suggestion['suggested_text']}\"\n"
   
   # Generate updated content
   updated_content = your_llm.generate(feedback_prompt + "\n\nOriginal:\n" + doc["content"])
   
   # Push the update
   requests.patch(
       f"{BASE}/documents/{doc_id}",
       headers=headers,
       json={"content": updated_content}
   )

5. Resolve processed feedback

   # Resolve comments that were addressed
   for comment in comments:
       requests.patch(
           f"{BASE}/documents/{doc_id}/comments/{comment['id']}",
           headers=headers,
           json={"resolved": True}
       )
   
   # Accept or reject suggestions
   for suggestion in suggestions:
       action = "accept" if should_accept(suggestion) else "reject"
       requests.patch(
           f"{BASE}/documents/{doc_id}/suggestions/{suggestion['id']}",
           headers=headers,
           json={"action": action}
       )

Using the CLI

The same workflow using the CLI:

# Push initial document
notebind push report.md

# Create share link
notebind share DOC_ID --permission comment

# Check for feedback
notebind comments DOC_ID
notebind suggestions DOC_ID

# After updating the file locally
notebind push report.md

# Resolve feedback
notebind resolve DOC_ID COMMENT_ID
notebind accept DOC_ID SUGGESTION_ID

Tips

Tip

Anchored comments are your friend. When creating feedback prompts for your LLM, include the anchor_text and character offsets. This tells the agent exactly which part of the document the reviewer is referring to.

• Don’t resolve feedback before processing it — use resolved/unresolved status to track what’s been addressed
• Accept suggestions programmatically — if a suggestion is a simple text fix, accept it via the API instead of manually rewriting
• Use pull for a complete snapshot — the CLI pull command returns the document with all comments and suggestions in one JSON blob
• Create separate share links per reviewer — this lets you revoke access per person if needed