HeyHomees Lab Notes · No. 11 · Anthropic · Published October 28, 2026
By HeyHomees Newsroom ·
In 30 seconds
- It audits existing subagents, then lets you merge, fold, or keep them as tools
- Each change runs a short evaluation and shows before-after metrics
- The kit runs offline tests without an API key and estimates token cost before paid runs
What is it?
The Single-Agent Loop Kit is a starter kit for moving a multi-agent system toward one agent loop that loads skills and tools on demand. It provides steps to audit, baseline, merge, fold, and maintain subagents, and includes example code and scripts.
How does it work?
The kit reads your orchestration code, audits each subagent, builds or fills an evaluation, runs a baseline, then iteratively applies changes (merge, fold, tool conversion) each followed by an evaluation run. It uses Claude Code commands for building evals, hill-climbing, prompt audits, and cost optimization.
What does it need to run?
You need a repository containing your agent code, Claude Code access for the commands, and a way to run your system on test requests. No API key is required for the offline tests, and the kit estimates token usage before any paid runs.
By the numbers
- tasks in offline stand-in run: 19
- harness errors in offline run: 0
- initial free check tasks: ten
- repetitions for initial runs: two or three
The honest read
The kit reports that an offline dry run processes 19 tasks with zero harness errors, but this run is not scored because the stand-in follows keyword rules; real scoring happens only on paid model runs. The only quantitative figures (19 tasks, 0 errors, ten initial tasks, two or three reps) are self-reported by the repository documentation and do not reflect performance metrics such as latency or cost. The evaluations and token estimates are generated by Claude Code sessions, so any cost or speed numbers would be specific to the user’s model choice and usage, not inherent to the kit itself.
What we don't know yet
- Exact token cost or latency of the evaluation runs
- How the kit performs on large-scale production systems
- Supported programming languages or frameworks beyond the generic orchestration code
Sources
- README, GitHub (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.