Our summary
A small decision and a multi-step investigation need different control flows. Avi Chawla explains a System One setup in which application code supplies a bounded question and uses the model's answer. A System Two setup lets a larger model propose successive steps as tool results change what it knows.
The distinction is not simply small model versus large model. Both setups leave tool permissions, spending limits and final execution with application code. HarnessRouter gives a product one way to start tasks, watch progress, share files and continue sessions across different coding services. Its small-decision path uses a separate controlling program.
A shared task interface does not make agents reason alike or guarantee their outputs. Self-hosting does not mean every model runs locally. The post links a longer setup article, but neither establishes a controlled accuracy or savings comparison. Cancelling work can stop future steps; it cannot undo an email already sent.
Key takeaways
- Use bounded judgments when the answers are known in advance, and a planning loop when the path must develop.
- Keep permissions, budgets and side effects under application control in both designs.
- Standardise task handling without pretending different agents have identical behavior or reversible actions.