Engineering
CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution
A new paper introduces CoTrace, a framework for training terminal (command-line) coding agents that explicitly matches training data to the…
Engineering
SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles
A new paper introduces SkillForge, a reinforcement-learning method for managing agent skill libraries through a fitness-driven lifecycle of trial, active,…
Engineering
Mechanics of Long-Context Hybrid Models: From Hybrid Attention to Hybrid Position
A new study examines why hybrid attention architectures, which combine full attention with either sliding-window attention or gated linear-attention variants,…
Engineering
WEFT: Scaling Tool-Use Post-Training for General-Purpose Agents
WEFT is a framework for scaling tool-use post-training of general-purpose AI agents by jointly coordinating environments, tasks, agent harnesses, and…
Engineering
VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks
VeriHarness turns a large language model into an agentic verifier equipped with workspace tools and reusable verification skills for checking…
Engineering
Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence Modeling
This paper introduces triadic linear attention, which extends linear-attention mechanisms by replacing matrix-valued recurrent states with three-dimensional tensor states. The…
Engineering
Dynamic Harness Search: Building Multi-Agent Systems Per-Query via Prediction
Researchers introduce SHIFT, a method that builds a custom agent harness for each query instead of relying on a single…
Engineering
Periscope: Extending Frozen Language Models Beyond Their Context Window
Researchers present Periscope, a training-free inference method that lets a frozen language model answer questions over text far longer than…
Engineering
Beyond Semantic Similarity: Performance and Costs of Agentic Retrieval for Complex Tasks
Researchers evaluate agentic retrieval, which combines large language model reasoning with a retriever in a ReAct-style loop, against standard dense…
Engineering
Raven: The Harness of Harnesses for Composable Agentic Intelligence
Raven is an open-source multi-agent ecosystem that automatically constructs and evolves modular ‘harnesses’ for specific models and domains, treating each…
Engineering
Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression
The paper identifies ‘phase sensitivity’ in transformers using chunked KV-cache compression for long-context inference: retrieval accuracy varies systematically (by up…
Engineering
TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces
TraceDance automatically builds targeted agent-behavior benchmarks from real deployment traces rather than fixed benchmark suites, using an Anchor-and-Confirm pipeline (programmable…