4 ms·

I built Claude sub-agents that keep my context clean

2 points by houseworth 1y ago

houseworth 1y ago
Hey HN, I've been thinking about the actual use-case for sub agents and how they could solve problems I run into doing day-to-day dev. I have 6 use cases but built three of them so far and have some good results testing so far. When Claude searches my codebase or analyzes a big git diff, my context window fills up with thousands of tokens of crap that the master instance almost never needs. After a couple of these operations, I'm basically starting over because there's no room left for actual work. I built specialized sub-agents that run in their own contexts and return condensed summaries instead of dumping everything into my main conversation. house-research: Searches codebases/docs and returns condensed findings with file:line references instead of dumping full file contents into your context. house-git: Analyzes diffs and commits, returns impact summaries grouped by severity instead of raw git output. house-bash: Runs commands and parses output for key info (pass/fail counts, relevant errors) instead of flooding your context with full logs. Three I haven't done yet: house-mcp (mcp), house-vision(screenshots), house-data(data queries) Results: 90-95% token reduction on heavy operations. My master instance stays focused on implementation instead of getting clogged with search results and logs. What operations eat your context that I should build agents for? One time install with CC: Clone https://github.com/houseworthe/house-agents https://github.com/houseworthe/house-agents. Detect my OS and copy the .claude directory to my current project. Verify all three agent files exist in .claude/agents/ directory. List available sub-agents to confirm they load. Test house-research by finding TODO comments.
bazhand 1y ago
I like this approach, and it's very similar to what I'm currently using - albeit my use case is to get Claude to automatically invoke the relevant subagents during both planning and build stages. Subagents using MCP would solve quite a lot of the context preparation required for a successful session. I'm using GLM-4.6 vision/web search MCP in claude code to relative success as the subagents can access these via API (not true MCP behaviour but close enough)
- houseworth 1y ago
  This is really cool. Really interested in your MCP workaround - I have three future agents blocked by Issue #7296. Would you be willing to share more details about your GLM-4.6 setup? Specifically how you're doing "API (not true MCP behaviour)" to get sub-agents accessing those tools? house-mcp would be super useful for me

2 points by houseworth 1y ago

Hey HN, I've been thinking about the actual use-case for sub agents and how they could solve problems I run into doing day-to-day dev. I have 6 use cases but built three of them so far and have some good results testing so far. When Claude searches my codebase or analyzes a big git diff, my context window fills up with thousands of tokens of crap that the master instance almost never needs. After a couple of these operations, I'm basically starting over because there's no room left for actual work. I built specialized sub-agents that run in their own contexts and return condensed summaries instead of dumping everything into my main conversation. house-research: Searches codebases/docs and returns condensed findings with file:line references instead of dumping full file contents into your context. house-git: Analyzes diffs and commits, returns impact summaries grouped by severity instead of raw git output. house-bash: Runs commands and parses output for key info (pass/fail counts, relevant errors) instead of flooding your context with full logs. Three I haven't done yet: house-mcp (mcp), house-vision(screenshots), house-data(data queries) Results: 90-95% token reduction on heavy operations. My master instance stays focused on implementation instead of getting clogged with search results and logs. What operations eat your context that I should build agents for? One time install with CC: Clone https://github.com/houseworthe/house-agents https://github.com/houseworthe/house-agents. Detect my OS and copy the .claude directory to my current project. Verify all three agent files exist in .claude/agents/ directory. List available sub-agents to confirm they load. Test house-research by finding TODO comments.

bazhand 1y ago

I like this approach, and it's very similar to what I'm currently using - albeit my use case is to get Claude to automatically invoke the relevant subagents during both planning and build stages. Subagents using MCP would solve quite a lot of the context preparation required for a successful session. I'm using GLM-4.6 vision/web search MCP in claude code to relative success as the subagents can access these via API (not true MCP behaviour but close enough)

houseworth 1y ago
This is really cool. Really interested in your MCP workaround - I have three future agents blocked by Issue #7296. Would you be willing to share more details about your GLM-4.6 setup? Specifically how you're doing "API (not true MCP behaviour)" to get sub-agents accessing those tools? house-mcp would be super useful for me