What's this about?
Multi-day autonomous software development with continual improvement — an outer harness that manages the inner harness across a run measured in days rather than hours. 11 upvotes.
Why does multi-day change the problem?
Because everything that was manageable over an hour breaks over a week. Context cannot be retained, so what survives compaction becomes the design question. The environment drifts — dependencies update, tests that passed start failing for unrelated reasons. Errors accumulate rather than being caught in a single review. And the agent's own earlier decisions become constraints it no longer remembers making. The "continual improvement" part is the interesting claim: the system is supposed to get better over the run rather than degrading.
Why does this matter to you?
- Multi-day autonomy is the capability that turns agents from tools into something closer to staff, and its failure modes are barely mapped.
- Third harness-evolution paper in three weeks — this is now a clear research front.
- It is also, uncomfortably, the exact configuration of the Hugging Face incident: agents running for days with degrading oversight.