LoopX is a local-first control plane for long-running agent work. Agent runtimes such as Codex, Claude Code and DeepSeek Harness execute each turn, and LoopX keeps the durable state between turns: the active goal, ordered todos, claims and leases, user gates, evidence, quota and recovery. It is not an agent framework or a provider-specific orchestration runtime, and it is not an autonomous production controller: dangerous permissions, publishing, production writes and final ownership stay with the human.
The 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, and the board is a projection while LoopX state remains the source of truth. Registered agents are peers: claims, leases, task boundaries, capabilities and typed continuation decide who acts next, with no durable leader identity required. An automatic turn checks quota first and records spend only after validated writeback, and a user gate that blocks one lane can let a separately audited safe fallback continue without bypassing the gate.
The architecture has four boundaries. The Kernel owns goal, todo, gate, evidence, quota, recovery and scheduling truth. A Capability defines a provider-neutral contract for one bounded, verifiable outcome, with shipped lanes for Issue Fix, Change Quality, Progress-Review Sentinel, Integration Branch, Explore, Decision Context and Periodic Report. A Provider calls an external system and returns observations, and an Extension packages an optional provider with an install, enable, upgrade and rollback lifecycle. Execution runs Agent to Capability to Provider, and control returns Provider readback to Capability transition to Kernel.
A Personal Agent Workspace (a browser or PWA app, with native desktop previews) shows what needs you, what is running, watched, scheduled or stopped, lets you steer a live turn or queue a message, and reviews protected changes through typed preview, explicit confirmation and receipts. Todos and gates can be projected into Lark Kanban, and a connected Lark conversation serves as an async inbox. Beyond Codex, Claude Code and DeepSeek Harness, host adapters cover KunlunCode, OpenCode, Pi, ZCode, Antigravity CLI and Kiro CLI, and any shell, scheduler or custom runner can drive the CLI directly. Optional Reward Memory (with OpenViking as one provider option) connects feedback to the next attempt, and Explore evaluates candidate changes against a baseline and guardrails.
On Long-Horizon Terminal-Bench (46 tasks, GPT-5.6 Sol at maximum reasoning effort, web search disabled), the project reports these results, with one effective trial per task and mode and unequal runtime budgets, so they are observed system results rather than an equal-budget efficiency estimate.
| 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%) |
Features
- Durable goal state: the active goal, scope, authority, ordered todos, claims, leases, gates, evidence and run history persist across sessions, restarts and harness changes
- Quota-aware continuation: each turn asks whether to deliver, ask, wait, self-repair or stay quiet, and spend is recorded only after validated writeback
- Peer agent teams: registered agents claim todos and leases before delivery and update them after validation, so ownership and evidence stay visible with no leader identity
- User gates: concrete decisions waiting on the owner instead of a vague waiting state, with separately audited safe fallbacks that never bypass a gate
- Capability lanes: Issue Fix, Change Quality, Progress-Review Sentinel, Integration Branch, Explore, Decision Context and Periodic Report, each reporting maturity, provider readiness and write boundaries
- Host adapters: Codex App and CLI, Claude Code, DeepSeek Harness, KunlunCode, OpenCode, Pi, ZCode, Antigravity CLI, Kiro CLI, plus a worker bridge and custom runner guide
- Personal Agent Workspace: a read-first browser or desktop view of attention, running work, schedules, deliverables, periodic reports and their evidence, with typed preview and confirmation for protected changes
- Lark projection: todos and gates projected into Lark Kanban and an async inbox from a connected Lark conversation, with LoopX state remaining authoritative
- Recurring presets: daily triage, changelog drafts and PR watching as read-only presets, with a readiness score for a connected recurring goal
- Review packet: a compact owner-facing view of decisions, evidence, validation and unresolved gates
- Reward Memory and Explore: optional, default-off paths that feed feedback into the next attempt and evaluate candidate changes against a baseline, treatment and guardrails
