Im building an agent that interacts with health care providers on your behalf to proactively get itemized bills and other substantiation as needed. Born out of my own frustration with the paperwork for getting reimbursed from my healthcare cost sharing ministry. The app is React with a Laravel backend on EC2 where the agent runs.
Recently Ive been experimenting more with coding from my phone using Claude Code for the Web. Its basically turned Github Actions into my development environment. Its enabled me to fire off a quick prompt in planning mode, go play with my kids, review and approve the plan while Im cooking dinner, then let it go 10 rounds with the AI code reviewer while I put the kids to bed. As a busy parent I feel way more productive than if I had to carve out sit down focus time for a side project.
Loved this writeup. I have built an agent for a specific niche use case for my clients (not a coding agent) but the principles are similar. ive only implemented 1-4 so far. Going to work on long term memory next, but I worry about prompt injection issues when allowing the LLM to write its own notes.
Since my agent works over email, the core agent loop only processes one message then hits the send_reply tool to craft a response. Then the next incoming email starts the loop again from scratch, only injecting the actual replies sent between user and agent. This naturally prunes the context preventing the long context window problem.
I also had a challenge deciding what context needs injecting into the initial prompt vs what to put into tools. Its a tradeoff between context bloat and cost of tool lookups which can get expensive paying per token. Theres also caching to consider here.
Totally agreed! Ive had good success using claude code with Cucumber, where I start with the spec and have claude iterate on the code. How does ossature compare to that approach?
Looking forward to when this gets on Bedrock. I built an app with a niche AI agent and to this point only Sonnet is really good enough for our use case, but its expensive!
Im in a slight disagreement with our CTO about the value of writing acceptance criteria yourself. When I write my own acceptance criteria its a useful tool forcing me to think through how the system ought to work. Definitely in agreement that writing is an important tool for clarifying thinking, not just generating context.
We hired on a software engineer fresh out of college this January, and we are a very claude heavy shop. We have noticed and have explained to him that we still have to do the real engineering work: requirements decomposition, interface design, verification and integration. Claude is helpful but he is still responsible for his output. He’s been doing well so far! Theres still plenty of things for him to bang his head on and grow.
Recently Ive been experimenting more with coding from my phone using Claude Code for the Web. Its basically turned Github Actions into my development environment. Its enabled me to fire off a quick prompt in planning mode, go play with my kids, review and approve the plan while Im cooking dinner, then let it go 10 rounds with the AI code reviewer while I put the kids to bed. As a busy parent I feel way more productive than if I had to carve out sit down focus time for a side project.
https://www.healthsharetech.com