Daily Learnings: Fri, Jun 20, 2025
Where there is great love, there are always miracles. — Willa Cather
Today’s experiments revolved mostly around continued tweaking of Claude Code and working with various MCP servers to see how it improved the development process.
Context7 MCP
I watched this YouTube video discussing the context7 service that is a free repository of documentation for LLM and AI-assisted coding agent consumption.
First of all, this service is awesome. Great to know that this exists. You can navigate to the site and copy / paste the library information directly into a file in your repo if you’d like (which, to be honest, I might just opt to do instead of using the MCP server). The MCP setup is quick, and can be found here, but interacting with the model to find the right docs and then load them into context could be a little more cumbersome than finding what you need and creating markdown files in your repo.
Of course, there might be autonomous use cases that would be better served with an MCP server, but for traditional software project, I think the copy/paste approach works.
Claude Code Pricing Updates
In my testing today, I switched to the $20/mo plan authentication and ran through some light use cases. So far it’s not hit any limits that I’ve seen, but today was a lighter coding day.
Experimenting with Smaller Agents
As I’ve worked with Claude Code this past week a bit more in-depth, I’ve pondered on what might help it perform better. This morning I was listening to an episode of the Changelog podcast in which they were interviewing Amanda Silver, CVP Microsoft DevDiv. They were discussing agents, and the topic came up about smaller, more narrowly-focused agents.
I think that one of the main issues I’ve been dealing with is that I’ve assumed Claude Code (or other agents) should be able to comprehend our entire codebase, the entire solution that we’re trying to build, and all context on the services that we’re connecting to, and that all code that’s generated will be consistent and follow our patterns. This is the standard that I’d hold a human developer to, and that I hold myself to.
Technically, I can do my best to give the LLMs backing the agents all of this context through priming and asking to review files. But even when things are loaded in context windows, I’ve found that inconsistencies show up.
So, I’m going to also experiment a bit with smaller agents that are coordinated by a synchronous workflow. Here are some nascent ideas on some of those agents, applied to a SOLVD repo that handles a lot of automation for our business. These agents assume the Github Issue-driven development model.
- All Agents
- All agents should be primed on our custom types and our conventions on using Pydantic for models
- Should all be expected to write tests for what they build
- Should be able to run the tests and get feedback on / iterate on what they build
- To start, they won’t commit what they’ve done, but maybe in the future
- Overall designer agent
- Takes Github issue and designs the entire solution, store it in MD file that’s checked into source control
- Will have the most knowledge of the codebase and our overall desires
- Primed most especially on workflows and related utils
- Utils agent
- Works off of the generated MD file and the original GH issue
- Focuses on the utils, integration classes, and the things that will be used at the orchestrator level
- Primed just on our current utils and their designs
- Orchestrator agent
- Works off of the generated MD file and the original GH issue
- Build the overall logic and workflows
- Primed mostly on our existing workflows for examples
- API layer agent
- Works off of the generated MD file and the original GH issue
- Builds the api endpoint & celery logic
- Primed on those for examples
- Final code review agent
- Works off of the generated MD file and the original GH issue
- Reviews all code to be committed
- Maybe this is already done with Claude Github Action?
I know that Claude Code and other tools like Roo can spin off “sub agents” to do tasks like these without having to do it manually. I’d like to eventually get there, but I’m interested in being more involved in the process to start, to ensure our process is right, and that the priming is giving appropriate context to the agents.