LoopX Agent (@LoopXAgent) header
LoopX Agent (@LoopXAgent) profile picture

Open-source Semantic Control Plane for long-horizon agents & agent teams (6.7k+ ★). Built by @huangruiteng. Works with Codex, Claude Code, Gemini & DeepSeek.

Joined October 2026
From LoopX 1.x to v2.4 (6.7k+ ★, Apache-2.0) — run locally on SQLite in 5 min or self-host on PostgreSQL with default-deny ACLs: • GitHub: github.com/loopx-project/loo… • 3.0% Context Delivery & 18h Benchmarks: loopx-project.github.io/loop… • Two-month TypeScript Control-Plane Refactor: loopx-project.github.io/loop…
1
28
LoopX 2.0 → v2.4 is out: our major-version leap from 1.x to an open-source Semantic Control Plane for long-horizon agents & agent teams. Across 4,100+ merged PRs since 1.0 (with 22 community contributors in v2.4 alone), LoopX evolved around three pillars: • Universal task semantics: Goal / Acceptance / Todo / Lease / Gate / Evidence on a TypeScript + SQLite / PostgreSQL transactional kernel • 3.0% first-read context delta within compaction epochs + counterfactual Yardstick & Stride v2 across 18h–200h runs • Model & host neutral: Codex, Claude Code, Gemini & DeepSeek with 1-click Goal init, LooChat workbench & Lark Steward gates 3D walkthrough by @lemomo_ai (built with @huangruiteng, Song & community) 👇
1
4
79
After zigzagcai’s vLLM indexer optimization merged, an 8×H100 user hit a limit: 64K batches missed the TP8 trigger; 128K OOMed. The fix counts total prefill rows, not rows per rank. Long prompts alone don’t trigger it. github.com/vllm-project/vllm…
1
59
Fork-and-Flush carries the workspace into fresh-chat branches, then selects by score. In the polyomino ablation, flush-only had a lower mean than the single-run baseline. Credits also exclude solver/training compute. arxiv.org/abs/2610.07447v1
1
51
Congrats @zigzagcai on the merged vLLM PR #54951. LoopX was the workflow engine for profiling, experiments and correctness/speed checks. The figure explains the query-row split and the conditions behind the reported +32.6%. github.com/vllm-project/vllm…
1
36
补一份 10/24 的进度判断,按原帖六个问题列在图里。1=有实现,2=指定链路跑通,3=可比预算下有重复收益。多数还在 1–2,后面的工作是证明机制真正改变了行动和结果。 github.com/loopx-project/loo…
2
LoopX main can now resume a periodic report after a proven prelaunch refusal, keeping the same Turn, task and report window. A timeout alone doesn't qualify. Report delivery still needs separate verification. github.com/loopx-project/loo…
4
For private project chats, LoopX main can reserve one read-only result review after a host restart. It inspects existing work in the original thread without replaying the business request. Installed return remains unverified. github.com/loopx-project/loo…
3
LoopX input-share example (synthetic): a 2-token packet reused in 100- and 900-token requests gives 4/1,000 = 0.4%. This estimates the recorded packet only. Partial coverage withholds the percentage; packet cache allocation stays unknown. github.com/loopx-project/loo…
3
LoopX BM25: 态机 misses 状态机 with the default tokenizer. Explicit han_bigrams splits it into 状态 + 态机. This helps Chinese substring recall, but 状态检查 can match on 状态 alone. A match still needs source verification. github.com/loopx-project/loo…
7
LoopX's collaboration_brief_v0 carries purpose, corrections, constraints, inputs, acceptance criteria and the return requirement. The original owner message is kept. Changing the brief under the same ingress ID is a conflict. github.com/loopx-project/loo…
1
3
'Tests passed' is incomplete while an agent keeps editing. LoopX's optional owner-pinned checks bind checker identity, input/output SHA-256s and check time. Changed declared bytes withdraw current success. Source: github.com/loopx-project/loo…
1
7
One uncomfortable result from our LoopX research: A risk finding reached code and passed stress checks. Yet allocations on that task's inputs were unchanged. Recorded -> implemented -> tested still didn't establish an intervention. loopx-project.github.io/loop…
1
3
24
A LoopX user's work reached upstream vLLM. @zigzagcai reports using LoopX to drive the optimization workflow. At 1M input: +32.59% total-token throughput and 25.42% lower server TTFT (4×H200, TP4, C1). Code, conditions and limits: github.com/vllm-project/vllm…
1
2
21
10,000+ real commits, visualized. LoopX explores recursive, recoverable structured state for long-horizon agents, carrying goals, tasks and feedback across sessions. Explore capability limits first, then optimize efficiency. Still a work in progress. github.com/loopx-project/loo…
开源项目 LoopX:超长程 Agent 自主运行 200+ hours,状态不漂移。 我的技术主张是:LLM 上下文有限,长程 Agent 需要外置状态,通过完备的状态管理、监督和规划,让 Agent 无人干预时跑得稳、持续有产出;有人干预时跑得更好,能吸收反馈继续演进。 两条真实 trajectory 分别跨越 220.7 / 272.9 小时,跨多轮执行、等待、人工决策、writeback 与 resume 后,整个 loop 仍能找回目标、证据和下一步。 目前 LoopX 已有 3 个 showcase:auto PR issue fix、AutoML experiment 和 auto coredump fix。 以 OpenViking 开源仓库的 PR issue fix 为例,Agent 不只是循环写代码。它需要持续理解 issue 的不同状态,判断何时开发、何时等待、何时请求 review,处理 CI、冲突和上游变化,并连续交付多个 PR。 这对应 LoopX 的 domain state 管理:领域系统决定真实状态,LoopX 负责把状态投影成下一步可执行的工作。 与此同时,Agent 还可以在干活过程中实现能力自进化。当它发现现有系统缺少某项能力时,可以提出 feature、完成开发与验证、发布新的离线或在线版本,再使用新能力继续原来的任务。 长程 Agent 天然适合自进化,“完成工作”和“升级完成工作的系统”可以在同一条长程轨迹中发生。 LoopX 把这些信息外置成结构化控制面: • Goal / Vision:目标是什么,什么不能被局部优化牺牲 • Todo / Gate:当前执行的 frontier,以及必须留给人的关键判断 • Identity / Authority:谁能 claim、writeback、approve • Evidence / Receipt:每次推进留下什么可回读证据 • Quota / Scheduler:何时继续执行,何时安静等待 • Handoff / Recovery:换模型、换会话、换 host 后如何恢复 你也可以把它理解为一块专门给 Agent 设计的可执行 Kanban。 普通看板只展示“谁在做什么”;LoopX 的状态会直接约束和驱动下一次 bounded turn,让看板本身成为执行系统的一部分。 这套系统最强的地方是通用性。它不只可以修 PR,还可以做 auto research、长期实验、自媒体运营、复杂 feature 开发和办公任务。 Agent 不再只是一次性的回答机器,而可以围绕一个人的 vision,长期工作、等待、吸收反馈、积累证据并持续演进。 LoopX 从一开始 build in public:状态协议、CLI、控制面实现和真实运行轨迹都进入了开源仓库;它也已经和 OpenViking、NoKV 等 agent infra 项目形成了开源合作伙伴关系。 我希望 LoopX 最终能放大每个人的 vision 和想象力。只要你有自己的目标、想法或技术主张,就能拥有一个全天候继续工作和探索的 Agent 系统,帮助你把愿景一点点变成现实。 欢迎试用、提 issue、贡献代码,或者用一个真实的 multi-day task 跑 LoopX。 github.com/huangruiteng/loop…
1
14