🤖Anthropic's Claude Agent SDK Tackles Long-Running Coding Agents Across Sessions

Your long-running coding agents just got a memory fix

Every new agent session starts with zero memory of the last one. That's the core problem with long-running coding agents, and Anthropic just published a practical fix using the Claude Agent SDK. Here's the catch: the SDK's compaction alone won't get an agent through an arbitrarily long job. Left alone, the agent tries to do too much at once, runs out of context mid-implementation, and a later instance looks around, sees partial progress, and declares the job done. If you've shipped agent-written code, you've probably seen that movie. The fix splits the work in two. An initializer agent runs once with a different prompt, creating an init.sh script, a claude-progress.txt file, a feature requirements file, and an initial git commit. Then coding agents make incremental progress, commit with descriptive messages, and leave a clean state for the next session. Anthropic also found that testing matters: Claude marks features done without checking unless told to use browser automation and test like a human user. Vision limits and browser tooling gaps still let some bugs slip through.