Let the agent learn. Don't let it ship what it learned unchecked
In preprints, a deployed shopping agent, a memory system and a lifelong-learning benchmark all point the same way: self-updates help only when a frozen candidate is audited against the old behaviour before it goes live.
- In a preprint, a governed self-training system called G-CAST is now deployed in Yuancai, a B2B procurement agent on Alibaba's 1688 platform. Self-reported m:…4759
- In a preprint, a governance framework for self-modifying agents separates generation from adoption, using fresh paired audits, non-inferiority checks and calibrated statistical bounds before bounded deployment. Self-reported m:…8830
- In a preprint, the MIRIX-2.0 memory system reaches 91.4% on LoCoMo and 72.6% on LongMemEval-S, leading its strongest baseline by 14.3 points. Self-reported m:…4758
- In a preprint benchmark, no harness-based method gains on unseen tasks, while weight-based learning shows substantial gains. Self-reported m:…5225
The agent that updates itself in production
A model trained once falls behind when demand changes. That is the premise of a preprint describing G-CAST, a governed continual self-training system that converts daily production evidence under evolving demand and capability contracts into reward-routed, entropy-prioritized training data. Demand drift alters the distribution, intent and complexity of user requests, the authors write, and harness evolution requires the model to use newly introduced capabilities correctly without losing behaviour that users already rely on. Periodic manual retraining is too delayed and labour-intensive for this setting, so the system learns continuously instead. It is not a lab demo: the authors say G-CAST is now deployed in Yuancai, a B2B procurement agent on Alibaba's 1688 platform. m:…4759
The key word is governed. G-CAST uses governed checkpoint transition under cumulative evaluation to control each model update, letting the agent learn newly introduced behaviours without sacrificing previously accepted capabilities. The authors report stronger adaptation to newly introduced demand while better preserving previously accepted behaviour compared with a strong autonomous post-training baseline, and say ablations show both the data orchestration and the governed reward evolution matter to those gains. The system is designed so that learning and shipping are not the same step. m:…4759
A separate gate for adopting changes
A second preprint makes the same separation explicit as a governance framework. Autonomous agents modify prompts, memories, tools, workflows and policies using interaction histories, its authors write, and that creates three problems: candidate selection is adaptive, evaluators may be imperfect, and local gains can hide degradation against an earlier baseline. Their framework, called evidence-governed adoption, separates candidate generation from confirmatory adoption, evaluates frozen candidates on fresh paired audits, and combines candidate-wise Hoeffding bounds with independent calibration samples comparing the operational evaluator with a prespecified reference procedure. m:…8830
Adoption requires meaningful improvement on a prespecified primary objective, protects declared capabilities through reference-relative non-inferiority tests, and applies governance constraints before bounded canary deployment. The structure is deliberately boring: nothing ships because it looked better in a chat. It ships because a frozen version passed a fresh audit that was designed before the candidate existed. That is the difference between an agent that learns and an agent that drifts. m:…8830
Memory with receipts, and learning that actually transfers
Memory is part of the same problem, and a third preprint offers a memory system built for audit. MIRIX-2.0 is a multi-agent memory system for LLM agents that the authors say achieves high accuracy and efficiency across question-answering and procedural benchmarks without fine-tuning, reaching 91.4% on LoCoMo and 72.6% on LongMemEval-S, where it leads the strongest baseline by 14.3 points. A provenance hypergraph indexes episodic and semantic records with frames that cite their rows, so a recalled fact can be traced back to its source, and a component called AutoDream merges duplicate anchors outside the write path without rewriting any record, and distils task sessions into skills. The system also reaches 49.2% on MetaClaw-Bench, against 44.0% for MetaClaw's skill library. m:…4758
A fourth preprint, Continual-ARC, tests whether agents actually learn over a working lifetime, using a lifelong stream of ARC-AGI tasks to compare memory-based and weight-based continual learning under frozen-weight agent harnesses. Its results are sobering for the harness approach: persistence in the harness generally does not improve memorization unless provided with explicit expert guidance, most harness-based methods still learn less than a floor that simply caches each task's demonstrations, and no harness-based method gains on unseen tasks, whereas weight-based learning shows substantial gains. One concrete result: a Qwen3.5-9B model trained with DAgger to imitate GLM-5.3 Flash spent 29% less than its untrained self on 100 unseen tasks. The lesson for an executive is that letting an agent accumulate notes is not the same as letting it learn, and the only updates worth shipping are the ones that survive a frozen audit against the old behaviour. m:…5225
What is still unresolved
- Whether governed self-training results from a single deployed shopping agent generalise to other domains such as coding or customer service.
- How expensive the frozen-candidate audit step is in practice, since none of the papers reports its cost.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.