Inversion of Control
Problem
Traditional "prompt-as-puppeteer" workflows force humans to spell out every step, limiting scale and creativity.
Solution
Give the agent tools + a high-level goal and let it decide the orchestration. Humans supply guard-rails (first 10 % + last 3 %) while the agent handles the middle 87 %.
Example (flow)
sequenceDiagram
Dev->>Agent: "Refactor UploadService to async"
Agent->>Repo: git grep "UploadService"
Agent->>Tools: edit_file, run_tests
Agent-->>Dev: PR with green CI
References
- Raising An Agent - Episode 1, "It's a big bird, it can catch its own food."