Agentic AI

EvolveTrade replaces static, hand-written tool-use policies in LLM trading agents with a dynamic, text-parameterized system prompt. A dedicated Policy Agent updates this prompt after each interval by analyzing accumulated decision traces and realized portfolio feedback. This approach allows the agent to adapt its evidence gathering, tool invocation, and risk management strategies without modifying the underlying backbone LLM.

• Replaces rigid, pre-deployment tool-use policies with adaptive, text-based system prompts.
• Leverages a separate Policy Agent to refine strategies based on historical decision traces.
• Uses realized portfolio feedback to drive continuous improvement in risk and signal verification.
• Maintains the base LLM architecture, focusing updates solely on the policy layer.
• Enables agents to adjust behavior dynamically in response to changing market regimes.

GraphEcho reveals that LLM agents often treat repeated traversal of the same graph paths as independent corroboration, leading to inflated confidence. The benchmark isolates this by varying path counts while keeping evidence content constant, showing that redundant walks skew judgment across frozen models. Provenance-aware post-training reduces these revisits and improves synthetic accuracy but narrows source diversity and hurts performance on scientific claims.

• LLMs conflate repeated graph path visits with new evidence, skewing judgment confidence.
• GraphEcho benchmark isolates structural redundancy from evidential content for testing.
• Provenance-aware post-training cuts revisits but reduces distinct source coverage.
• Accuracy gains from PAPT on synthetic data do not generalize to scientific claims.

The HarnessTax study quantifies the friction introduced by AI coding assistants, focusing on the time and cognitive load required to integrate their output into existing codebases. It highlights that while agents generate code quickly, the subsequent review, debugging, and refactoring steps often negate those gains. The research suggests that the true cost of using these tools is not just in their subscription price, but in the structural changes they impose on development velocity.

• AI agents introduce significant review and debugging overhead that reduces net productivity gains.
• The 'tax' includes cognitive load from understanding generated code rather than writing it.
• Integration friction often outweighs the speed of initial code generation.
• Teams should measure end-to-end delivery time, not just code creation speed.
• Workflow adjustments are necessary to mitigate the structural impact of agent-generated code.

Cloudflare has open-sourced a coding-agent skill that transforms AI assistants into structured security auditors. The tool orchestrates isolated agents through six phases, including reconnaissance, coverage-led hunting, and independent verification of findings. It serves as the foundational starting point for the multi-stage, fleet-wide vulnerability discovery harness previously described in Cloudflare engineering blogs.

• Uses a six-phase workflow: recon, hunting, validation, output, verification, and reporting.
• Generates machine-readable findings with independent record verification for accuracy.
• Serves as the single-repo entry point for Cloudflare's internal vulnerability harness.
• Maps architecture and trust boundaries to identify deterministic coverage gaps.

A controlled study across six Qwen3 models (0.6B-32B parameters) compares supervised fine-tuning (SFT) with LoRA, reinforcement learning via GRPO, and a combined SFT-then-GRPO approach. SFT with LoRA proves to be the strongest method for in-distribution performance, winning 15 out of 18 experimental settings. While GRPO shows slight advantages in cross-dataset transfer scenarios, its margin over SFT remains under one point, suggesting SFT is the more robust baseline for tool-calling agents.

• SFT with LoRA is the top performer for in-distribution tool-calling across 0.6B-32B models.
• GRPO offers marginal gains in cross-dataset transfer, averaging less than one point over SFT.
• Combined SFT-then-GRPO does not clearly surpass standalone SFT for in-distribution tasks.
• Practitioners should prioritize SFT with LoRA as the primary adaptation method for tool use.

This paper introduces ERPBench, a new benchmark designed to evaluate screenshot-only computer-use agents on a live, reproducible ERP system rather than general desktop tasks. It addresses the unique challenges of enterprise software, such as dense interfaces and multi-step interactions that can alter persistent business records. Unlike existing benchmarks that rely on proprietary platforms or simulated approximations, this approach scores tasks against grounded state changes.

• Targets computer-use agents operating via screenshots in complex enterprise environments
• Evaluates agents on live, reproducible ERP systems instead of simulated approximations
• Focuses on dense interfaces and multi-step interactions typical of finance and procurement
• Scores performance against grounded state changes to catch errors affecting business records
• Moves beyond general desktop/web benchmarks to address specific enterprise software challenges

This research addresses the challenge of distributed perception in multi-agent Vision-Language Model systems, where different agents inspect distinct visual regions. It introduces a framework for shared visual context that enables agents to build on each other's contributions, recover missing context, and reconcile conflicting interpretations as new evidence emerges. The work focuses on memory hierarchy and consistency mechanisms to synchronize dependent reasoning across agents.

• Solves the distributed perception problem in multi-agent VLM setups
• Enables agents to share visual context across different image regions
• Provides mechanisms to reconcile conflicting agent interpretations
• Supports recovery of missing visual context during complex tasks