feat : slight improvement for planner & web agent

prompts/base/coder_agent.txt

@@ -42,7 +42,7 @@ Some rules:
 - You have full access granted to user system.
 - Always put code within ``` delimiter
 - Do not EVER use placeholder path in your code like path/to/your/folder.
-- Do not ever ask to replace a path, use current sys path or work directory.
+- Do not ever ask to replace a path, use work directory.
 - Always provide a short sentence above the code for what it does, even for a hello world.
 - Be efficient, no need to explain your code, unless asked.
 - You do not ever need to use bash to execute code.

prompts/base/planner_agent.txt

@@ -2,7 +2,7 @@ You are a project manager.
 Your goal is to divide and conquer the task using the following agents:
 - Coder: A programming agent, can code in python, bash, C and golang.
 - File: An agent for finding, reading or operating with files.
-- Web: An agent that can conduct web search, wrapped with selenium it can interact with any webpage.
+- Web: An agent that can conduct web search and navigate to any webpage.
 
 Agents are other AI that obey your instructions.
 

prompts/jarvis/planner_agent.txt

@@ -1,8 +1,8 @@
 You are a planner agent.
 Your goal is to divide and conquer the task using the following agents:
-- Coder: An expert coder agent.
-- File: An expert agent for finding files.
-- Web: An expert agent for web search.
+- Coder: A programming agent, can code in python, bash, C and golang.
+- File: An agent for finding, reading or operating with files.
+- Web: An agent that can conduct web search and navigate to any webpage.
 
 Agents are other AI that obey your instructions.
 

sources/agents/browser_agent.py

@@ -70,7 +70,7 @@ class BrowserAgent(Agent):
         {search_choice}
         Your goal is to find accurate and complete information to satisfy the user’s request.
         User request: {user_prompt}
-        To proceed, choose a relevant link from the search results. Announce your choice by saying: "I want to navigate to <link>"
+        To proceed, choose a relevant link from the search results. Announce your choice by saying: "I will navigate to <link>"
         Do not eplain your choice.
         """
     
@@ -104,7 +104,7 @@ class BrowserAgent(Agent):
           - If it does and you completed user request, say REQUEST_EXIT.
           - If it doesn’t, say: Error: <why page don't help> then go back or navigate to another link.
         2. **Navigate to a link by either: **
-          - Saying I want to navigate to <url>: (write down the full URL, e.g., www.example.com/cats).
+          - Saying I will navigate to <url>: (write down the full URL, e.g., www.example.com/cats).
           - Going back: If no link seems helpful, say: GO_BACK.
         3. **Fill forms on the page:**
           - Fill form only on relevant page with given informations. You might use form to conduct search on a page.
@@ -149,6 +149,8 @@ class BrowserAgent(Agent):
         You previously took these notes:
         {notes}
         Do not Step-by-Step explanation. Write Notes or Error as a long paragraph followed by your action.
+        You might REQUEST_EXIT if no more link are useful.
+        Do not navigate to AI tools or search engine. Only navigate to tool if asked.
         """
     
     def llm_decide(self, prompt: str, show_reasoning: bool = False) -> Tuple[str, str]:
@@ -200,7 +202,7 @@ class BrowserAgent(Agent):
             if "note" in line.lower():
                 saving = True
             if saving:
-                buffer.append(line)
+                buffer.append(line.replace("notes:", ''))
             else:
                 links.extend(self.extract_links(line))
         self.notes.append('. '.join(buffer).strip())
@@ -214,7 +216,7 @@ class BrowserAgent(Agent):
         return None
     
     def conclude_prompt(self, user_query: str) -> str:
-        annotated_notes = [f"{i+1}: {note.lower().replace('note:', '')}" for i, note in enumerate(self.notes)]
+        annotated_notes = [f"{i+1}: {note.lower()}" for i, note in enumerate(self.notes)]
         search_note = '\n'.join(annotated_notes)
         pretty_print(f"AI notes:\n{search_note}", color="success")
         return f"""
@@ -307,6 +309,7 @@ class BrowserAgent(Agent):
 
             links = self.parse_answer(answer)
             link = self.select_link(links)
+            self.search_history.append(link)
 
             if "REQUEST_EXIT" in answer:
                 pretty_print(f"Agent requested exit.", color="status")
@@ -334,7 +337,6 @@ class BrowserAgent(Agent):
             if speech_module: speech_module.speak(f"Navigating to {link}")
             self.browser.go_to(link)
             self.current_page = link
-            self.search_history.append(link)
             page_text = self.browser.get_text()
             self.navigable_links = self.browser.get_navigable()
             prompt = self.make_navigation_prompt(user_prompt, page_text)

sources/agents/code_agent.py

@@ -35,7 +35,7 @@ class CoderAgent(Agent):
         info = f"System Info:\n" \
                f"OS: {platform.system()} {platform.release()}\n" \
                f"Python Version: {platform.python_version()}\n" \
-               f"\nYou must work in directory: {self.work_dir}"
+               f"\nYou must save file in work directory: {self.work_dir}"
         return f"{prompt}\n\n{info}"
 
     def process(self, prompt, speech_module) -> str:

sources/agents/planner_agent.py

@@ -81,8 +81,9 @@ class PlannerAgent(Agent):
     def show_plan(self, json_plan: dict) -> None:
         agents_tasks = self.parse_agent_tasks(json_plan)
         if agents_tasks == (None, None):
+            pretty_print("Failed to make a plan.", color="failure")
             return
-        pretty_print("▂▘ P L A N ▝▂", color="status")
+        pretty_print("\n▂▘ P L A N ▝▂", color="status")
         for task_name, task in agents_tasks:
             pretty_print(f"{task['agent']} -> {task['task']}", color="info")
         pretty_print("▔▗ E N D ▖▔", color="status")
@@ -94,7 +95,10 @@ class PlannerAgent(Agent):
             animate_thinking("Thinking...", color="status")
             self.memory.push('user', prompt)
             answer, _ = self.llm_request()
-            pretty_print(answer.split('\n')[0], color="output")
+            for line in answer.split('\n'):
+                if "```json" in line:
+                    break
+            pretty_print(line, color="output")
             self.show_plan(answer)
             ok_str = input("Is the plan ok? (y/n): ")
             if ok_str == 'y':
@@ -122,7 +126,7 @@ class PlannerAgent(Agent):
         agents_tasks = self.parse_agent_tasks(answer)
 
         if agents_tasks == (None, None):
-            return "Failed to parse the tasks", reasoning
+            return "Failed to parse the tasks.", reasoning
         for task_name, task in agents_tasks:
             pretty_print(f"I will {task_name}.", color="info")
             pretty_print(f"Assigned agent {task['agent']} to {task_name}", color="info")
@@ -135,7 +139,4 @@ class PlannerAgent(Agent):
             except Exception as e:
                 raise e
             agents_work_result[task['id']] = self.last_answer
-        return self.last_answer, ""
-
-if __name__ == "__main__":
-    pass
+        return self.last_answer, ""

sources/router.py

@@ -344,8 +344,8 @@ class AgentRouter:
             return "HIGH"
         elif complexity == "LOW":
             return "LOW"
-        pretty_print(f"Failed to estimate the complexity of the text. Confidence: {confidence}", color="failure")
-        return None
+        pretty_print(f"Failed to estimate the complexity of the text.", color="failure")
+        return "LOW"
     
     def find_planner_agent(self) -> Agent:
         """