Repos Worth Reading

Part III reads open-source projects in the agentic-coding space at a named commit, so every claim can be checked against the exact code it describes. These are readings, not product reviews: each chapter says what the code and the documentation show on the day it was read, and marks what it could not verify.

Start with SWE-agent and the agent-computer interface. It explains the 2024 argument that a model needs an interface built for it, and why the same group's documentation now points newcomers to a successor of about a hundred lines. Aider is read for two ideas that are now everywhere, a ranked, token-budgeted map of the repository and strict edit formats, and ships a standard-library repo-map script. OpenHands at v1.16.0 shows a larger system: the agent step loop, workspaces and sandboxes, the condenser and the confirmation gate. ccusage is a different kind of project, a local cost dashboard, read for where its dollar figures come from, with a script that re-runs its rules.

Read these alongside Harnesses, which covers the commercial harnesses from their documentation. Every chapter gives its commit, and a project may have moved on since. No further reading is queued for this part at the moment; the plan's open item for the book is the comparison update that waits for a harness's next major version.