Task Decomposition and Explicit Task Lists

A repair shop does not rely on a mechanic's memory for a large job. When a car needs new brakes, new tires, and an oil change, the shop prints a work order with each job on its own line. The mechanic signs a line only after that job's test has passed, and a mechanic who arrives after lunch can read the sheet and continue.

An agent on a long task needs this even more. The model keeps nothing between calls, so what it "remembers" about the remaining work is only what it can find in a long history.

The Running Case

Forge is a coding agent that works on a repository