notes, projects, and tools from the lab. everything runs on free tiers.
multi-agent orchestration built entirely on free LLM providers. two-ceo model: human decides what and why, LLM decides how. sub-agents never spawn sub-agents.
view on githubmay 2026
don't pay for what you can get free. the whole project runs on free llm providers — groq, gemini, github models, cloudflare, and local ollama. paid apis are a last resort. the cost ladder works.
two-ceo model. the human decides what and why. the llm decides how. the llm orchestrates sub-agents, but sub-agents never spawn sub-agents — that is a hard structural rule.
context is everything. sessions cap at ~80k tokens, then hand off. three strategies: compaction (summarize, discard redundant output), external memory (files as persistent state), and sub-agent delegation (child burns its own context, parent only sees a summary).
match rigor to scope. a one-session project doesn't need the full 9-agent pipeline. sometimes it is just researcher -> developer -> documenter.
ask when unsure. one clarifying question saves five fix messages. guessing burns context and time.
fresh context per task. commit after each task. if something fails, retry only the failed subtask, not everything.
sub-agents cost ~2-3k tokens. only worth it when you need isolation, a different model, or restart tolerance.