Agentic AI

The AgentKVShift paper introduces a training-free method to reduce inference latency in LLM agents by reusing Key-Value (KV) caches for structured agentic memory. Unlike prior techniques designed for raw RAG passages, this approach uses probe-guided residual correction to handle metadata like summaries and tags. The method selectively recomputes only necessary tokens per retrieved memory unit, addressing the high prefill costs associated with re-encoding structured context.

• Reduces prefill latency by reusing KV states for structured agentic memory units
• Uses probe-guided residual correction to maintain accuracy without retraining
• Outperforms existing RAG-focused KV reuse methods on metadata-heavy contexts
• Eliminates full re-encoding of retrieved summaries, keywords, and tags

LeafData is an agentic system designed to replace manual, error-prone JSON configuration in modern data migration pipelines. It features a frontend chatbot that incrementally gathers user intent and performs schema-driven validation, passing validated inputs to a backend service. The backend then generates executable JSON configuration artifacts directly consumable by orchestration platforms, enabling end-to-end pipeline generation.

• Eliminates manual JSON authoring for data connections and pipeline logic
• Chatbot interface captures user intent with incremental schema validation
• Backend generates validated, executable configs for orchestration platforms
• Reduces time and errors associated with domain-specific migration setup

This paper proposes a technical definition for AI-native systems, moving beyond marketing hype by focusing on decision-level authority rather than model capability. It distinguishes between occupancy (who executes a decision) and revision authority (who can change it), organizing the latter into a ladder of self-tuning, self-rewriting, and self-architecting. A system is classified as AI-native only when an AI autonomously rewrites its own implementation code.

• Shift focus from model capability to 'revision authority' for precise AI-nativeness classification.
• Distinguish 'occupancy' (execution) from 'revision authority' (modification rights) in system design.
• AI-nativeness requires the AI to autonomously rewrite its own system implementation code.
• Ladder of authority includes self-tuning, self-rewriting, and self-architecting capabilities.
• Paper addresses ambiguity in 'AI-native' by providing a formal decision-level model.

Researchers introduce a longitudinal evaluation instrument that generates synthetic conversations from seeded fact scripts rather than extracting keys from existing chats. This approach embeds validity intervals and volatility classes into the data, ensuring gold answers are constructed mechanically and verified for answerability. The resulting corpus addresses documented label errors and contamination issues found in standard short-history benchmarks.

• Inverts benchmark pipeline: seeds facts first, then renders synthetic chat from manifests.
• Eliminates label contamination by mechanically instantiating questions from valid scripts.
• Introduces volatility classes and validity intervals missing in standard short-history tests.
• Corpus includes ~380 questions across 15 types to stress-test long-term memory retention.

T3 Code provides a lightweight web-based interface to manage and interact with various coding agents, including Codex, Claude, Cursor, and OpenCode. Users must authenticate with at least one provider via their respective CLI tools before utilizing the GUI. The tool supports immediate execution via npx or installation as a desktop application across Windows, macOS, and Arch Linux.

• Consolidates multiple AI coding agents into a single minimal web GUI interface.
• Requires pre-authentication via CLI for each supported provider (e.g., codex login).
• Accessible instantly via npx t3@latest without local installation.
• Available as desktop apps for Windows (winget), macOS (Homebrew), and Arch Linux (AUR).

FlowEvo is a training-free framework that captures successful execution traces from LLM agents and compiles them into persistent skill records. Each record pairs a callable artifact with structured guidance, subject to interface, replay, and safety checks. These skills are stored in an inference-time bank to systematically benefit future tasks rather than remaining transient.

• Eliminates need for retraining by persisting successful agent workflows at inference time
• Pairs callable code artifacts with structured guidance for reliable reuse
• Applies safety and replay checks before admitting new skills to the bank
• Targets the gap where useful procedures are lost after solving a single task

Standard RAG pipelines break chronological context into isolated snippets, losing critical temporal structure for decision-making. This paper introduces TLM, an agentic framework that uses SHAP-guided feedback to iteratively refine evidence sets. It employs a latent growth curve model over chunk embeddings to detect trends, turning points, and information gaps in long-form text.

• RAG pipelines discard temporal structure, hurting chronological decision accuracy
• TLM uses closed-loop agents to iteratively refine retrieved evidence
• SHAP-guided feedback directs the retrieval refinement process
• Latent growth curve models detect trends and gaps in text embeddings

This paper identifies Role Drift, a failure mode where compound LLM modules abandon their assigned responsibilities while maintaining overall task accuracy through hidden shortcuts. The authors propose Role Anchor, a regularization technique that constrains training by monitoring how role prompts shift next-token predictions relative to neutral prompts. This approach makes internal role violations observable and correctable during end-to-end reinforcement learning.

• End-to-end RL optimizes for output accuracy, not internal role adherence, enabling silent failure modes.
• Role Drift allows modules to use role-violating shortcuts that bypass system-level evaluation metrics.
• Role Anchor regularizer tracks prediction shifts caused by role prompts to enforce behavioral constraints.
• Using neutral prompts as a baseline provides a proxy for intended role effects during training.