AgentDescent is a research framework from developer Danyang Chen that reframes self-improving AI agents through a gradient-descent metaphor of the authors' own coining: instead of one agent revising itself serially, N worker agents propose parallel diffs against a shared, version-controlled artifact -- a prompt, a skill file, a directory of files, or a rule list -- and a single aggregator thread runs a five-stage merge pipeline (selection, conflict resolution, fusion, acceptance, promotion) that commits accepted changes to a git-backed ledger, with a lag-budget-bounded asynchronous mode for running barrier-free. Which strategy backs the artifact decides whether concurrent edits can fuse at all: a SingleSlot prompt always contradicts, while a FileTree or KeyedRules artifact lets edits on different files or categories merge without conflict. A companion paper (DOI via Zenodo) backs the README's measured claims: on a one-key artifact where naive keyed-union merging fuses 0 of 48 proposals, reflective merging fuses 42 of 48, cutting model calls by roughly 40% at a fixed rollout budget and reaching a median 6.8x wall-clock speedup over a serial baseline on a GEPA/HotpotQA setup. AgentDescent also reimplements 19 published self-evolution algorithms as plug-ins under one shared scheduler -- eight reproduced against their original benchmarks (ACE, GEPA, DGM among them) and eleven looser analogues (PromptBreeder, Reflexion, Voyager among them) -- turning the scheduler into a controlled variable across papers rather than a property baked into each one's own loop. It also ships as an agent-side plugin: an MCP server and CLI let Claude Code, Codex, OpenCode, and DeepSeek Harness sessions start, monitor, and apply a run, falling back to that host's own CLI when MCP sampling is unavailable. Rollouts run on the host machine or inside a resource-limited container provider. The authors call this a research reference implementation rather than a production system, and their own docs flag that resuming interrupted stragglers, reusing discarded evidence cards, and cluster-aware scheduling across artifacts have scaffolding but are not yet implemented.