Agentic AI

This paper addresses the challenge of language-model agents performing tasks that span days or weeks, exceeding standard context windows and human attention spans. The authors propose that agents must run continually without forgetting, a capability achieved through an external harness rather than model architecture changes. The proposed solution is a hierarchical system featuring time-indexed levels, bounded file summaries, and a clocked tick mechanism for autonomous action.

• Agents need a continuous, non-forgetting harness to handle long-term tasks beyond context windows.
• Architecture uses time-scale levels where each keeps a bounded summary of the level below.
• Clocked ticks serve as the fundamental unit for autonomous agent actions.
• Cascaded intelligence manages complexity across these hierarchical levels.

This paper addresses the challenge of AI agents operating in enterprise systems where business rules are dynamic and undocumented. The authors introduce a method for continual world model discovery, allowing agents to infer causal relationships by observing outcomes of their actions on records. Evaluation uses EnterpriseWorldShift, a benchmark built on a live ServiceNow environment with nine tables and 25 hidden rules.

• Agents must learn business logic dynamically rather than relying on static schemas.
• World models are updated continuously as organizational rules change over time.
• Benchmark uses live ServiceNow data to test rule discovery accuracy.
• Focuses on causal inference from record interactions in complex enterprise workflows.

This paper challenges the assumption that RL gains in agentic language models reflect better decision-making, arguing that endpoint success conflates arrival state with execution capability. Because agents in closed loops generate their own observation sequences, SFT and RL checkpoints are evaluated from fundamentally different state distributions. The authors introduce checkpoint handoffs to isolate these effects, revealing that restricting comparisons to shared states can misleadingly flip the perceived impact of RL training.

• RL gains may reflect state selection rather than improved policy execution.
• Closed-loop agents create dependent observations, biasing standard evaluations.
• Comparing SFT vs RL on identical tasks is flawed due to divergent state paths.
• Checkpoint handoffs isolate arrival state from in-state decision quality.
• Selection bias in restricted state comparisons can invert performance signals.

LangChain 1.4.1 is a patch release addressing two specific issues in the 1.4.0 series. It resolves a bug where open Model Context Protocol (MCP) object arguments were not being preserved correctly. Additionally, the release corrects documentation for the InterruptOnConfig feature.

• Updates are required if you rely on open MCP objects to ensure arguments are passed correctly.
• Review InterruptOnConfig usage against the corrected documentation to avoid configuration errors.
• This is a minor patch release focused on stability and documentation accuracy.
• Upgrade from 1.4.0 to ensure proper behavior with MCP integrations.

A new position paper argues that the fragmentation in agentic AI stacks mirrors the pre-OS era of computing. It proposes a Foundation Model Operating System (FMOS) layer to virtualize interactions with foundation models. This abstraction aims to provide portable state, memory, and guardrails, eliminating the need for each framework to re-implement these core services.

• Current agentic frameworks embed implicit runtimes, making behavior non-portable across tools.
• FMOS would abstract FM interactions like VMs abstract hardware for standardized access.
• Centralizing governance and state management could resolve brittle control planes.
• Protocols like MCP and A2A handle connectivity but not the underlying runtime abstraction.