LoopX · Blog
Design, engineering, and field notes on long-horizon agents.
Context delivery · Mechanisms and examples
· Updated November 3
One voice task explains JSON Pointers, batched selection, index depth, complete requirements, exact sharing and compaction-cycle deltas, paired with ANN 18h (7,159 vs 5,753 QPS) layered byte/token ablations and early-window/cross-task counterexamples.
Read articleResearch notes · EdgeBench and LHTB
Exploratory notes by songoow on historical, partly interrupted EdgeBench/LHTB runs. The feedback-first option is withdrawn; these are not official LoopX results or current capability claims.
Read the articleLoopX 2.0 · Engineering
From an incremental migration proposal to complete transactions, shared authority, semantic checks, and retirement of old paths. How real callers and counterexamples changed the design.
Read articleAgentic Software · Development
Using LoopX to explore infrastructure for agentic software: preserve goals, evidence, and human feedback across sessions, and carry exploration into verifiable delivery.
Read articleResearch notes · EdgeBench
· Revised November 3
Follow the research from early October to early November: reliable feedback, goals, experiment size and recovery, tested across New Foundations Consistency’s 18h plateau breakthrough, Schema’s plateau counterexample and five-arm causal attribution.
Read articleDesign perspective · Self-improvement
From its first Meta Goal to Personal Workspace: four months of bootstrapping, 6,848 non-merge commits, and a case for making improvements compound across rounds.
Read the articlePostmortem · Field notes
Our own advisory agent answered from stale, partial, or unread sources five times while every health signal stayed green. What failed, why local fixes were not enough, and the contract that makes the gap visible.
Read the articleScenarios · Practice and research
See how goals, evidence, and delivery responsibility work across three scenarios, alongside signals from SWE-Marathon, DeepSWE × Sol, and V4 Flash max.
Read the articleAgent-native Kanban · Long-running collaboration
Four diagrams explain Goals, Todos, claims, leases, gates, evidence, and shared authority state—and how they support long-running collaboration.
Read the articleArchitecture · Design notes
How LoopX uses goals, a state kernel, and Effect Programs to carry work across sessions, preserve human judgment, and recover from interruptions.
Read the articleBy 张彦峰ZYF · Original Chinese article on CSDN · CC BY-SA 4.0.
A community perspective on coordinating long-running agents through goals, evidence, human decisions, and recovery. Its bounded verified transitions (BVT) and proposed metrics are the author's design suggestions; assess current LoopX behavior against the Developer Book and implementation evidence.
The article cites ArshTechPro's introduction to LoopX among its sources. This reading entry links to the author's original text.
Read the original on CSDN