agentsmd
Agents.md is missing Agents.json - a system configuration file
When I handle a task, I want to choose the right model for it.
The current solution only proposes instructions. I want to tailor temperature, MCPs, maybe code scaffolding for examples and running the results in a preconfigured environment.
That's how I did it here: https://github.com/teabranch/adm-workbench/tree/main/.agent
This lets you split the task into smaller tasks, and no matter what solution you provide, it will always be better when you tailor the environment to it. It also reduces irrelevant context and static environment for repeatable tasks.
Are you suggesting a subset of MCPs to be used for a particular kind of task?
How would the agent evalue the kind of task it's asked to do, and condition itself accordingly?
Consider agent each agent is configured for a single kind of task.
It can be generating a class, an app (CLI?), web page.
The configuration defines for it the MCPs (figma, playwright..), what model best fits it (maybe a a smaller model can handle it or a fine tuned one for that task), temperature and more.
Now you can define agents for specific tasks, and operate them to generate types of code. (A class that fits a specific project or company projects)
The agent creates the code, tests it, and commits to a branch.
If you handle this agentically, you can orchestrate these agents from your main agent.