Why this matters
Long-running AI work breaks assumptions that single-session agents rely on: objectives drift, evidence stales, ownership changes, and schedulers can keep spending after useful work ends. LoopX solves this by separating a compact, durable control-state from any specific runtime so bounded agent turns can resume, hand off, and be audited without granting agents unchecked production permissions.
What Sets It Apart
- Durable, read-first kernel: stores goal state, typed todos, claims/leases, gates, and compact run history as the authoritative source rather than chat transcripts or ephemeral scheduler state — so progress and decisions remain legible across days.
- Quota-aware, gate-first cadence: quota checks, concrete user gates, and safe fallbacks decide whether an agent turn should run, ask for judgment, or stay quiet, reducing wasted cycles and accidental actions.
- Runtime-agnostic integrations: adapters and prompts for Codex, Claude Code, Cursor, shell and custom runners let existing agent toolchains execute bounded turns while LoopX manages continuation and handoff.
- Evidence-backed handoffs and review surfaces: todos carry identity, authority, and evidence;
review-packetand dashboard projections preserve decision lineage for human operators.
Who It's For and Trade-offs
Great fit if you run multi-day engineering, research, benchmark, or experiment objectives that need persistent scope, reviewer-facing evidence, and safe handoffs between agents or human owners. It also suits recurring monitoring, PR/issue workflows that must preserve review state, and multi-agent proposer/executor/evaluator patterns.
Look elsewhere if you need a full agent runtime, autonomous production controllers, or out-of-the-box hosted dashboards; LoopX is a local, control-plane kernel that deliberately avoids granting production credentials or performing unverified writes. Expect some setup to integrate host runtimes and an operational model that emphasizes read-first state and human-in-the-loop decisions.