AgentDescent
A parallel, asynchronous framework that lets several agent workers propose edits to a shared skill, prompt, or harness at once, merging their diffs the way gradient descent merges gradients.
AgentDescent treats agents as the parameters being optimized: multiple workers propose edits to a shared, version-controlled artifact in parallel, and a barrier-free aggregator merges the resulting diffs into one library through conflict resolution, fusion, statistical acceptance, and a transactional commit, rather than by averaging. This lifts the usual limit of accepting only one self-improvement change per iteration.
The project reports that on a benchmark where edits are confined to a single key, so a keyed union can never fuse changes, its reflective merge succeeds on 42 of 48 merge attempts versus 0 of 48 for a keyed union, at 40% fewer model calls and a median 6.8x wall-clock speedup over a faithful serial control on a GEPA/HotpotQA benchmark.
The design separates a fixed data path, where workers roll out tasks against ledger snapshots and an aggregator commits winning diffs, from a governance layer that is deliberately not configurable, so a self-modifying system cannot replace its own evaluator.