loopx

<div align="center"> <h1 align="center">LoopX</h1>

Give your agents a goal. Keep the work moving.

The open, local-first control plane for long-horizon agents and personal agent teams.<br> _{Keep goals, decisions and evidence across sessions. Work with Codex, Claude Code, DeepSeek Harness and other supported runtimes.}

Get started · Workspace · LHTB results · Docs · 简体中文

LHTB · 46 tasks · GPT-5.6 Sol: LoopX 1.0.3 Heartbeat reaches 0.4948 mean reward — +17.3% vs Plain Codex, +10.6% vs native Codex Goal.<br> _{<a href="#lhtb-results">Results and pass rates ↓</a>}

</div>

---

More verified work. Less human attention. LoopX gives agents durable goals, bounded continuation, peer ownership and recoverable handoffs. Your runtime provides the model and tools; LoopX keeps track of what to do next, what is accepted, and when to ask you.

<a id="learn-loopx"></a>

| What you want to do | Start here | | --- | --- | | Keep a coding or research agent working across sessions | Install and connect | | Manage personal projects, schedules and decisions in one place | Personal Agent Workspace | | Let agents collaborate and deliver verifiable results | Agent collaboration guide |

Meet the Personal Agent Workspace

Keep long-horizon goals in one local-first workspace. Goals, attention, conversations, tasks, files, schedules, and recovery stay durable across days, restarts, and harnesses. Reopen a project, inspect the previous turn’s state and evidence, and continue the next permitted action.

<a href="docs/assets/personal-workspace/loopx-dashboard-launch.mp4"> <img src="docs/assets/personal-workspace/workspace-1.0.webp" alt="LoopX Workspace: owner decisions, Agent tasks, scheduled watches and completed work" width="960"> </a>

LoopX 1.0 brings these long-horizon control states into the Personal Workspace. It gives you one place to:

• see what needs you, what is running, what is being watched, and what is scheduled or stopped;
• configure Goal capabilities, distinguish machine defaults from Goal overrides, and preview changes before applying them;
• steer a live turn, queue a message, or use the async inbox from a connected Lark conversation with explicit Goal/Agent/session routing;
• inspect deliverable files, periodic reports, and their supporting evidence;
• continue across Codex, Claude Code, direct-model, and other registered Agent sessions without losing Goal state or evidence;
• review protected changes through typed preview, explicit confirmation, and receipts while LoopX state—not the browser—remains authoritative.

For Manager group conversations, LoopX keeps message visibility separate from Turn authority; see the bilingual Lark Manager context and authority contract.

loopx dashboard

loopx dashboard is the supported browser/PWA launch path. You can also download native desktop previews from the 1.0 release; they reuse the same loopback services and Goal state. Apple Silicon macOS supports signed App updates that pair the shell with its bundled runtime, plus repair and recovery. Python 3.11+ is required; the App is ad-hoc signed, not notarized. Windows preview installers currently use manual updates and a separately installed CLI. Desktop installation, updates, and source development.

<details> <summary>Capability settings and reproducible workspace scenarios</summary> <img src="docs/assets/personal-workspace/capability-1.0.webp" alt="Real Workspace recording: configure child-task capacity and allowed responsibility domains" width="960">

From a source checkout, run python -m demo.workspace serve to explore a community event, a home-energy comparison, and a neighborhood website release. Each has four work roles, 18 tasks, two decisions, and two watches. The screenshot above comes from this reproducible workspace. Scenarios and replay instructions.

</details>

Watch the full 32-second walkthrough · Read the workspace guide · Try the five-minute tour

<a id="how-it-works"></a>

Why LoopX

An agent can finish a task in one session. Long-running work is harder: objectives change, owner decisions appear, evidence goes stale, agents hand work to peers, and a scheduler can keep spending after no useful transition remains. Chat memory and a timer are not enough to govern that.

LoopX keeps the durable control state in one compact layer:

objective / issue / project
   │
   ▼
LoopX state: objective + gates + todos + scope + evidence + quota
   │
   ├─ human judgment needed? ── yes ─▶ ask a concrete question and wait
   │
   ├─ safe fallback available? ──────▶ run one bounded agent slice
   │
   ▼
Codex / Claude Code / Cursor / shell agent executes one turn
   │
   ▼
write evidence + handoff + next todo ─▶ quota decides the next tick

Agent runtimes execute the work. LoopX governs the state that lets engineering, research, discovery, and operations loops continue across runs. It is not another agent framework or a provider-specific orchestration runtime.

A useful mental model is an agent-native Kanban for long-running work. Cards carry identity, authority, evidence, and continuation. Moves are validated operators such as claim, gate, monitor, and writeback. The board is a projection; LoopX state remains the source of truth.

Registered agents are peers. Claims, leases, task boundaries, capabilities, and typed continuation decide who acts next; no durable leader identity is required.

LoopX is useful when you run:

• multi-day engineering, research, benchmark, or experiment objectives;
• issue and PR loops that must preserve scope, evidence, and review state;
• recurring heartbeat or monitor work;
• projects with owner, safety, publication, or private-data gates;
• peer-agent teams where ownership, leases, and handoff matter;
• creator, research, or operations workflows whose progress must remain legible to a non-engineering operator.

LoopX is not an autonomous production controller. Dangerous permissions, publishing, production writes, and final ownership stay with the human.

Personal Agents, Teams, and Self-Improving Workflows

Meta Muse and Grok Bot make persistent personal agents and delegated work a familiar product idea. LoopX approaches that space as an open, provider-neutral control plane for agents you already run—not as a hosted replacement for either product.

• Personal agent: use the shipped Workspace and connected Lark surfaces to inspect goals, steer work and resolve decisions. The persistent steward and semantic handoff RFC extends this toward one capable front door for a team; the complete journey remains under qualification.
• Agent team: share goals without erasing worker ownership. Claims, evidence and explicit returns keep implementation and review connected. See the three-agent collaboration demo for a bounded example, including corrections and two review rounds.
• Self-improving workflows: connect feedback to the next attempt through Reward Memory, and evaluate candidate changes through Explore. Both are optional. The link to recursive self-improvement (RSI) is the engineering loop—propose, test, review, retain—not a claim that LoopX has demonstrated autonomous model improvement or compounding capability gains.

A personal steward is an interaction role, not a second source of authority. Background execution still needs an available host and configured runtime. The overall roadmap separates shipped foundations from the next end-to-end acceptance milestones.

<a id="see-it-in-action"></a>

Evidence

<a id="lhtb-results"></a>

LHTB: Higher Mean Reward With the Same Model

LoopX 1.0.3 Heartbeat reaches 0.4948 mean reward across 46 matched tasks: +17.3% over Plain Codex and +10.6% over native Codex Goal. All three use GPT-5.6 Sol, max reasoning effort, and disabled Web Search. Long-Horizon Terminal-Bench goes beyond coding: these tasks span research reproduction, scientific simulation, multimodal analysis, professional workflows, games, and systems work.

| Execution mode | Mean reward ↑ | Strict pass rate (≥0.95) | Pass rate (≥0.80) | | --- | ---: | ---: | ---: | | Plain Codex | 0.4218 | 7/46 (15.2%) | 12/46 (26.1%) | | Native Codex Goal | 0.4475 | 4/46 (8.7%) | 14/46 (30.4%) | | LoopX 1.0.3 Heartbeat | 0.4948 | 7/46 (15.2%) | 15/46 (32.6%) |

The ≥0.80 threshold is a supplementary, post-hoc view; ≥0.95 remains the benchmark's strict solved threshold. No score in this study equals 0.80, so these counts also match the study brief's >0.80 view.

Against native Goal, task-level rewards improve on 23, tie on 13, and regress on 10 tasks. Strict solves match Plain Codex, so the mean gain is not a claim that every task improves or that more tasks are fully solved than Plain. Mean reward also credits partial progress.

One effective trial per task and mode; designated replacement trials and unequal runtime/budgets, including longer budgets for some Heartbeat replacements. These are observed system results, not an equal-budget efficiency result or an isolated causal estimate of LoopX's effect.

Interactive results, task comparisons and methodology · Five-arm study and limits · Public task-level scores

Beyond benchmarks, LoopX also has inspectable long-running project evidence. The public OpenViking contribution sequence and the redacted, owner-run Auto ML showcase each span 200+ hours of elapsed loop lifetime, preserving bounded turns, decisions, and evidence updates. This measures wall-clock project time, not continuous model execution or unattended production autonomy. Open each visual to inspect the public-safe task graph, evidence branches, and cross-turn decisions; each case states its source and reproducibility boundary.

<details> <summary>Earlier 200+ hour project trajectories: OpenViking and ML experiments</summary>

Open-Source Issue Fix

200+ hour public contribution arc: PR delivery and reusable fix knowledge evolve together.

<a href="docs/assets/long-running-loop-openviking-trajectory.png"> <img src="docs/assets/long-running-loop-openviking-trajectory.png" alt="Open-source issue-fix trajectory linking focused PR delivery with reusable LoopX capabilities" width="760"> </a>