In my experience building these things over the years, the biggest mistake people make is trying to use those all-in-one autonomous wrappers like AutoGPT. They look cool in a demo but break way too easily for a real thesis. For a research agent on a tight budget, skip the hype and focus on a few core pillars. First, you need to understand state management—basically how the agent remembers past steps without getting stuck in a loop. Second, get comfortable with ChromaDB Open Source Vector Store for your memory. Its free and handles papers easily. Third, master structured output. If your LLM doesnt return clean JSON, your agent will crash. Stick to this lean stack for your three-week timeline:

• Python for logic
• LangGraph State Machine Library to control the flow
• OpenAI API GPT-4o-mini for cheap, fast reasoning
• Ollama Llama 3 8B for local testing Having tried many frameworks over the years, I know keeping it simple with a state machine is always better than chasing ghosts.

Regarding what #1 said about AutoGPT being a hype wrapper, i totally agree because i once spent a whole weekend trying to get a research bot to summarize papers only for it to get stuck in a loop and burn through my credits. It was a mess. If you are on a zero dollar budget, you definitely need to look at local hosting to avoid costs and keep things stable. It works okay if you keep expectations low and focus on reliability over fancy features. Here are the core technical bits i found actually matter:

• Loop control: You need a hard limit on iterations. I learned this the hard way when my agent started talking to itself until it crashed. Real autonomy needs a kill switch logic in your python code.
• Local LLMs: Since you have no budget, try running something like Mistral 7B v0.1 Instruct using Ollama Desktop for Mac/Windows. It runs locally so no API fees, which is a decent option for a thesis.
• Evaluation: You need to build a judge step. Basically, have one prompt do the work and a second prompt check if the work is actually what you asked for. It is slower but way more reliable.
• JSON parsing: Agents are useless if they output weird conversational text. Master using Pydantic or just regex to force the AI to give you structured data you can actually use in your next step. Building these things is mostly just sophisticated error handling disguised as AI. If you dont plan for the LLM to fail, it will fail at the worst possible time. Honestly, just keep it simple and dont chase the newest shiny framework tho.