loopx Wants to Be the Kernel Under Your Agent Team
A repo called loopx (github.com/huangruiteng/loopx) gained 327 stars today and passed 2,000 total. The pitch: a "loop engineering state kernel" for long-running AI agent teams. Strip the jargon and it is the missing ops layer for multi-day agent work — MIT licensed, Python, v0.4 and early but usable.
The design puts a persistent control plane between human judgment and agent execution. Objectives, decision gates, executable todos, evidence logs, resource quotas, and handoffs all live in durable state, while Codex, Claude Code, Cursor, or your own runtime executes bounded turns against it. The cycle: loopx checks quota and gates, an agent claims a task through a lease, completes the bounded segment, evidence gets recorded, and the kernel decides what happens next — proceed, wait, or stop and ask a human a concrete question. Gates require explicit decisions, not the vague "waiting for feedback" state where most agent workflows silently die.
Why this matters: everyone who has run agents for more than a day discovers the same thing — the model is fine, but the state is chaos. Which objective are we on, what did we already try, who decided X, how much have we spent. LongHorizon-Harness measured this from the research side this week: giving a fixed model external task-state added 30 points on long-horizon benchmarks. loopx is the same insight arriving from the practitioner side, with quotas so agents cannot spend unbounded money and claims so two agents cannot grab the same task. The agent-team kernel is becoming its own layer — whoever standardizes it holds the whole stack in place.
← Back to all articles
The design puts a persistent control plane between human judgment and agent execution. Objectives, decision gates, executable todos, evidence logs, resource quotas, and handoffs all live in durable state, while Codex, Claude Code, Cursor, or your own runtime executes bounded turns against it. The cycle: loopx checks quota and gates, an agent claims a task through a lease, completes the bounded segment, evidence gets recorded, and the kernel decides what happens next — proceed, wait, or stop and ask a human a concrete question. Gates require explicit decisions, not the vague "waiting for feedback" state where most agent workflows silently die.
Why this matters: everyone who has run agents for more than a day discovers the same thing — the model is fine, but the state is chaos. Which objective are we on, what did we already try, who decided X, how much have we spent. LongHorizon-Harness measured this from the research side this week: giving a fixed model external task-state added 30 points on long-horizon benchmarks. loopx is the same insight arriving from the practitioner side, with quotas so agents cannot spend unbounded money and claims so two agents cannot grab the same task. The agent-team kernel is becoming its own layer — whoever standardizes it holds the whole stack in place.
Comments