Lab note #071 Agentic loop for buying decisions

Switched gears this week to start a new, but related project. I started a prototype on an app that helps users make buying decisions with an AI agent. Lately, I've been using ChatGPT to help me make buying decisions, but I found some aspect of the experience wanting. So I started an experiment to see what other experience I could have. I mean, lots of people crying about how chat shouldn't be the final interface for AI, right?

It was pretty quick and dirty, as it was just FastAPI backend with HTML5 frontend (no framework!). I also did some manual editing to get things to work and learn some pretty obvious-in-hindsight things, just from doing this project this week.

1. Agentic loop can plan things out and do things when you give them the tools to do so.
2. They're stateless by nature, so are completely reliant on the message history you send them, including the results of function tool calls.
3. The effectiveness of agents depend largely on the context you can give it. So message history, RAG, and even the state of the application is all very helpful.
4. There is merit to a canvas based approach with tooling.

There's also some realizations with how to change my workflow to adapt to this new way of working. I figured that'd be something that would be interesting enough as its own blog post.

The other ah-ha moment for me this week was that this has very good alignment on a computational notebook. All the work that I've done up to this point, (CRDTs, Prolly Trees, Reactivity, Algebraic Effects, and DBSP) have culminated into an opinionated idea of what I think a computational environment should look like. And surprisingly this also include something like an app to make buying decisions.

Anyway, here's a lot of the agent bookmarks I saved for the past two years and am catching up on, if you care to follow along.

AI Agentic computing reading list

• What is an AI agent?
• How to Build an Agent - Amp
• Building Effective AI Agents \ Anthropic
• How we built our multi-agent research system \ Anthropic
• Is there a Half-Life for the Success Rates of AI Agents? — Toby Ord
• a-practical-guide-to-building-agents.pdf
• Cognition | Don’t Build Multi-Agents
• Building AI Products—Part I: Back-end Architecture
• The Agentic Systems Series - The Agentic Systems Series
• Design Patterns for Securing LLM Agents against Prompt Injections - 2506.08837v2.pdf
• Reverse Engineering Cursor's LLM Client · TensorZero
• ART·E: How We Built an Email Research Agent That Beats o3 | OpenPipe
• The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT
• Poking Around Claude Code
• ai agents are local first clients • Solving the decision problem
• Develop AI Agents for System Engineering in Factorio - 2502.01492v1.pdf
• aperoc/toolkami: Minimal AI agent framework that just works with only seven tools.