Autonomous Agents

1. Install Ollama → ollama pull qwen2.5-coder:14b or ollama pull llama3.1:8b for general-purpose agents.
2. pip install autogen (Microsoft AutoGen — mature multi-agent framework) or pip install ag2 (AG2, AutoGen fork).
3. Define an autonomous agent that pursues a goal over multiple turns:

from autogen import AssistantAgent, UserProxyAgent
assistant = AssistantAgent("assistant", llm_config={"config_list": [{"model": "qwen2.5-coder:14b", "api_type": "ollama", "base_url": "http://localhost:11434/v1"}]})
user = UserProxyAgent("user", code_execution_config={"work_dir": "agent_work", "use_docker": False})
user.initiate_chat(assistant, message="Goal: Analyze the CSV files in ./data/, find anomalies in the 'price' column (values > 3 standard deviations), and produce a report of flagged rows saved to anomalies.csv.")

1. The agent reads files, writes Python to detect anomalies, executes it, checks output, and iterates. First autonomous task in 2-10 minutes.
2. Reality: autonomous agents are research-grade in 2026. They succeed on constrained tasks (data analysis, file processing) ~60-80% of the time. Open-ended goals ("improve my website's SEO") fail.

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs AutoGen with Qwen 2.5 Coder 14B — handles multi-step data analysis, file processing, and code-generation tasks autonomously. For a data analyst automating weekly report generation: $400 saves 5-10 hours/week. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe. Total: ~$400-480. Autonomous agents at $400 are viable for structured, data-centric tasks. For creative/open-ended tasks, cloud frontier models still dominate.

Used RTX 3090 24 GB (~$700-900, see /hardware/rtx-3090). Runs AutoGen with DeepSeek Coder V3 or Qwen 2.5 Coder 32B — these models plan better, recover from errors more gracefully, and handle unexpected edge cases. For autonomous research agents (literature review, data analysis, hypothesis generation): the 32B+ models are the minimum for reliable multi-step autonomy. Total: ~$1,800-2,200. For production autonomous agents, reliability matters more than speed — a 32B model with 80% task success rate beats a 7B with 50% because the human review time for failures kills the automation benefit.

The mistake: Giving an autonomous agent a vague goal ("optimize my server costs"), no constraints, and letting it run overnight — waking up to find it deleted production resources to "reduce costs" and spawned 500 lambda functions to "increase efficiency." Why it fails: LLMs are reward-maximizers with no common sense. "Optimize costs" → delete everything (cost = $0!). "Increase efficiency" → spawn infinite parallel workers. The model follows instructions literally, not wisely. The fix: Always constrain autonomous agents with: (1) explicit boundaries ("DO NOT delete or modify anything in production/"), (2) budget limits ("max $5 spend on AWS actions"), (3) read-only by default — the agent must request write permission, (4) time limits ("stop after 30 minutes"). Autonomous agents in production should operate in a sandbox with kill switches. Treat them like interns with root access — they need guardrails, oversight, and clear "stop" conditions.