让 Agent
自己变强。Make agents
improve themselves.
AgentDescent 是一个并行、异步的自进化框架。定义要进化的对象，运行任务，按 held-out 结果合并每一处改进。AgentDescent is a parallel, asynchronous framework for self-evolving agents. Define the artifact, run tasks, and merge only changes that improve held-out evidence.
curl -fsSL https://raw.githubusercontent.com/Birfy/agentdescent/main/scripts/install.sh | bash定义。探索。
只合并更好的。Define. Explore.
Merge only better.
一个 evolve() 入口，战略决定改什么，Worker 并行提出候选，聚合器用 held-out 证据决定是否合并。策略槽位可以替换，循环保持不变。One evolve() entry point. A strategy decides what changes, workers propose in parallel, and the aggregator merges only candidates that improve held-out evidence. Swap one policy slot; keep the loop.
从“能进化”
到“能放心跑”。From “it evolves”
to “you can run it”.
这个预览包含仍在 review 的采纳、安全与可观测性工作：候选在评估前先过守卫，决策可记录，部署可灰度或停止，归档趋势可诊断。The preview includes adoption, safety and observability work still under review upstream: guards run before evaluation, decisions are recorded, deployments can be canaried or stopped, and archive trends can be diagnosed.
把答案挡在外面Keep the answer key out
在候选购买 held-out 评估前，检查 diff 是否泄露 gold、task id 或私有语料。Screen a candidate diff for gold answers, task ids or private corpus strings before it buys a held-out sweep.
LeakageGuard ↗ 02 / APPROVAL高风险变更要人批Require a human for risk
高风险自修改进入 JSONL 队列，审批绑定到真正被 review 的内容。Queue high-risk self-modifications and bind approval to the exact content that was reviewed.
ApprovalGate ↗ 03 / RUNTIME一个受控会话One guarded session
组合、注入验收、检查 kill switch、记录决策，一次调用完成。Compose, inject acceptance, honour a pre-flight kill switch and record the decision in one call.
run_guarded ↗ 04 / DEPLOYMENT灰度、回滚、停止Canary, roll back, stop
成对符号检验给出 promote / hold / rollback，文件开关停止下一次会话。A paired sign test recommends promote/hold/rollback; a file switch stops the next session.
canary · KillSwitch ↗ 05 / DIAGNOSIS说清为什么没学会Name the failure mode
把最差 rollout、被拒候选和 outcome 直方图变成一个结论与下一步。Turn the worst rollouts, refused candidates and outcome histogram into one verdict and the next step.
diagnose ↗ 06 / OBSERVABILITY盯住采纳率与成本Watch acceptance and cost
把 decision records 聚合成采纳率、每次采纳成本与 Prometheus 导出。Aggregate decision records into acceptance rate, cost per accepted change and a Prometheus export.
rsi_metrics ↗决策日志、统计证据、边际收益停止、风险策略和算法锦标赛也都已进入导航。Decision records, statistical evidence, marginal-gain stopping, risk policy and algorithm tournaments are also in the navigation.
从决策日志开始Start with the decision log ↗四个命令，
一次可复现的进化。Four commands,
one reproducible evolution.
先 plan 看成本，再 start 后台运行；完成后 show 看差异，确认后 apply。每一步都能在 status 和决策记录里追溯。Plan the cost, start the run in the background, inspect the diff with show, then apply after approval. Every step is traceable in status and the decision record.
文档是英文的；这个预览页默认中文，可切换到 EN。The docs are in English. This preview page defaults to Chinese; switch to EN at the top.
打开快速开始Open the quickstart ↗$ agentdescent plan spec.json
kind: skill_dir tasks: 18 workers: 4 calls <= 216
$ agentdescent start spec.json --detach
run_id: 7f3a1c status: running
$ agentdescent status 7f3a1c
round 3/6 held-out 0.71 committed 1 rejected 2
$ agentdescent show 7f3a1c --json
outcomes: {committed: 1, below-threshold: 2}
decision: adopted · evidence: held-out gain同一个循环，
换不同的搜索。One loop,
many search strategies.
ACE、GEPA、DGM、OpenEvolve 等 23 个方法端口都运行在同一个引擎上。它们展示如何替换 strategy、policy 或 aggregator，而不是限制你能做什么。ACE, GEPA, DGM, OpenEvolve and 19 more method ports run on the same engine. They show how to replace a strategy, policy or aggregator — not what you are allowed to build.
查看全部 23 个方法Explore all 23 methods ↗从一个 Agent 开始。Start with your agent.
一条命令安装 AgentDescent，连接 PATH 上支持的 Agent。One command installs AgentDescent and connects supported agents on your PATH.