Maintain agent goals and tasks across sessions with a control plane
When agents work on tasks over several days, you may have experienced issues where direction gets lost or evidence disappears once a session ends. LoopX is a tool that records what an agent should do next, what evidence exists, and when human confirmation is needed. It operates on top of existing runtimes like Codex or Claude Code, helping work progress steadily without human intervention.
Workflows that continue across sessions
LoopX manages agent tasks by grouping them into 'goals'. Each goal stores a to-do list, permissions, evidence, and conditions for moving to the next step. This state persists even if the agent ends a session or hands off to another agent. LoopX's state, rather than browser or chat history, becomes the single source of truth. This allows the agent to verify results from previous turns and perform only permitted next actions.
Agent teams and personal workspaces
LoopX provides a personal agent workspace that consolidates goals, schedules, conversations, files, and recovery states in one place. When multiple agents collaborate, ownership and leasing clearly distinguish who performs which task. Handoffs and verification between peer agents are possible even without a leader. Connecting to messengers like Lark allows you to issue task instructions or make decisions remotely. All changes are recorded through explicit confirmations and receipts, with LoopX state holding final authority rather than the browser screen.
Measured performance and real-world cases
LoopX's performance can be verified through LHTB benchmarks and real project cases.
- LHTB benchmark results Using the GPT-5.6 Sol model across 46 tasks, LoopX 1.0.3 Heartbeat recorded an average reward of 0.4948. This is 17.3% higher than standard Codex and 10.6% higher than native Codex Goal.
- Long-term project evidence Published OpenViking contribution cases and Auto ML experiments show elapsed times of over 200 hours each, demonstrating preserved limited turns, decisions, and evidence updates.
- User-reported cases Independent users reported achieving over 13 hours of C++ accuracy runs, 4 days of unattended execution, and 7 merged PRs through LoopX.
Installation and getting started
Python 3.11 or higher and Node.js 22.22.3 or higher (24 LTS recommended) are required. Install it with the command python3 -m pip install --upgrade loopx && loopx workflow-skills --install. After restarting the Codex App, enter $loopx <complex task> in a project thread, and LoopX will connect the project, plan the task, and set up heartbeats. In Claude Code, use /loopx <complex task>. You can check the current goal and pending decisions with the loopx status command.
Pre-use considerations
LoopX is not an autonomous production environment controller. Dangerous permissions, publishing, production environment writes, and final ownership remain with humans. Agents require an available host and a configured runtime to perform tasks. Default usage statistics are disclosed upon first use and can be disabled in settings. On Apple Silicon macOS, signed app updates are supported but are not notarized. The Windows preview installer uses a separately installed CLI for manual updates.