| """ | |
| Autonomous agent runner – uses AgentSession, AgentMemory, AgentAction. | |
| This is called by Celery and runs a multi‑cycle research loop. | |
| """ | |
| import json | |
| from .models import AgentAction, AgentMemory, AgentSession | |
| from .orchestrator import get_llm_response # we'll add a simple LLM caller | |
| from .tools import ToolRegistry | |
| def run_agent_cycle(session: AgentSession, user_id: int) -> dict: | |
| """Execute a single cycle of the autonomous agent.""" | |
| # 1. Get memory | |
| memory = AgentMemory.objects.get_or_create(session=session)[0] | |
| # 2. Build prompt from memory + objective + tool definitions | |
| tool_defs = ToolRegistry.get_definitions() | |
| system = ( | |
| f"You are an autonomous ML research agent. Objective: {session.objective}\n" | |
| ) | |
| system += f"Previous summary: {memory.long_term_summary or 'None'}\n" | |
| system += "Working context: " + (memory.working_memory or "None") | |
| messages = [ | |
| {"role": "system", "content": system}, | |
| {"role": "user", "content": "Proceed with the next step. You may call tools."}, | |
| ] | |
| # 3. Call LLM | |
| response = get_llm_response(messages, tools=tool_defs) | |
| # 4. Parse tool calls / actions | |
| for action in _parse_actions(response): | |
| tool_name = action["tool"] | |
| params = action["params"] | |
| result = ToolRegistry.execute(tool_name, params, user_id=user_id) | |
| AgentAction.objects.create( | |
| session=session, | |
| cycle=session.cycles_completed + 1, | |
| action_type="tool_call", | |
| tool_name=tool_name, | |
| input_data=params, | |
| output_summary=json.dumps(result)[:2000], | |
| success=result.get("status") != "error", | |
| ) | |
| # 5. Update memory | |
| if response.get("content"): | |
| summary = response["content"][:3000] | |
| memory.working_memory = summary | |
| if len(memory.long_term_summary or "") + len(summary) > 2000: | |
| # compress long-term summary | |
| memory.long_term_summary = (memory.long_term_summary or "")[ | |
| -1000: | |
| ] + summary | |
| else: | |
| memory.long_term_summary = (memory.long_term_summary or "") + summary | |
| memory.save() | |
| session.cycles_completed += 1 | |
| session.save() | |
| return {"cycle": session.cycles_completed, "status": "running"} | |
| def _parse_actions(response): | |
| """Extract tool calls from the LLM response (OpenAI function‑call format).""" | |
| actions = [] | |
| # simplified parsing – adapt to actual LLM output format | |
| if "tool_calls" in response: | |
| for call in response["tool_calls"]: | |
| actions.append( | |
| { | |
| "tool": call["function"]["name"], | |
| "params": json.loads(call["function"]["arguments"]), | |
| } | |
| ) | |
| return actions | |
Xet Storage Details
- Size:
- 2.8 kB
- Xet hash:
- 30d8a5acb527f70aaed9716ca17b7a4a95d7c0145082c80f4425d0aaecf39387
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.