Continual Learning for Long-Running Agents: Agents That Keep Getting Better

Value of the Information & Strength of the Argument

The talk provides valuable insights into the practical challenges of long-running agents and proposes a concrete solution (recursive language models) that is gaining traction. The argumentation is coherent, building from the problem of context rot to the solution of RLMs, and is supported by examples from industry (Perplexity, Claude Code) and benchmarks. However, the talk is also a promotional pitch for Prime Intellect, and the evidence for the effectiveness of RLMs is largely anecdotal or based on the speaker’s own claims. The case study with Ramp Labs is presented without detailed methodology, and the superiority of the trained model is not independently verified. Overall, the argument is persuasive but not rigorously substantiated.

Scientific Rigor, Source Quality, Title Accuracy

The talk references benchmarks (MRCR, GraphWalks) and industry tools (Claude Code, Perplexity) but does not provide specific citations or URLs. The description includes links to NVIDIA NeMo and Nemotron, which are relevant to the content. The title accurately reflects the content, which focuses on continual learning for long-running agents. The talk is an expert opinion rather than a peer-reviewed study, and the lack of detailed sources limits its scientific rigor. The speaker’s affiliation with Prime Intellect introduces a potential conflict of interest, as the talk promotes their services.

The title accurately reflects the content, which focuses on continual learning for long-running agents and how they can improve over time.

7/10

The talk presents a coherent argument for continual learning and recursive language models, supported by references to benchmarks (MRCR, GraphWalks) and industry examples. However, it is largely an opinion piece promoting the speaker's company, with limited peer-reviewed evidence and some unverifiable claims.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 01:00Introduction: Jack Min Ong introduces himself and the topic of continual learning for long-running agents.
• 03:00Discussion of the problem of long-running agents: context rot and performance degradation with long contexts.
• 05:00Introduction of recursive language models (RLMs) as a solution, with examples of programmatic tool calls and subagent delegation.
• 08:00Comparison of RLMs to chain-of-thought reasoning, arguing that RLMs are the next step in agentic AI.
• 10:00Explanation of the continual learning loop: harvesting trajectories, getting feedback, and retraining models.
• 12:00Presentation of Prime Intellect's platform: model training as a service, environments hub, evaluations, and sandboxes.
• 14:00Integration with NVIDIA stack: NeMo, Dynamo, and GPU/CPU sandboxes.
• 16:00Case study with Ramp Labs: training a smaller model that outperforms a larger model on a specific task.
• 18:00Discussion of data vendors using the platform to validate their data for training.
• 20:00Conclusion and call to action: visit primeintellect.ai.

• NVIDIA NeMo — Mentioned as part of the NVIDIA stack used by Prime Intellect for training and inference.
• NVIDIA Nemotron Models — Mentioned as supported models in Prime Intellect's platform.

• Recursive Language Models — The concept of recursive language models is central to the talk, and this paper provides the foundational research.
• GraphWalks Benchmark — The benchmark mentioned in the talk for evaluating long-context reasoning, supporting the claims about performance degradation.

• Context Rot — While the talk mentions context rot, this paper provides a more detailed analysis and may offer alternative perspectives on mitigation strategies.