Agentic AI

obra/superpowers introduces a structured development methodology for coding agents using composable skills. It enforces a workflow where agents analyze tasks before generating code, reducing hasty implementation. The framework integrates with multiple CLI tools like Cursor, Claude Code, and GitHub Copilot.

• Provides a methodology for coding agents via composable skills.
• Prevents hasty coding by enforcing initial analysis steps.
• Integrates with Cursor, Claude Code, Copilot, and others.
• Structured approach improves agent reliability and workflow.

The ai-memory project provides long-term memory solutions for AI coding CLI agents, allowing seamless transitions between different vendors like Claude Code and OpenAI Codex. It preserves context such as architecture details, failed approaches, and open questions when switching tools. The tool supports Linux, macOS, and Windows via WSL2, with Docker images and native binaries available.

• Maintains context across different AI coding agents to avoid re-explaining architecture
• Supports handoffs between vendors like Claude Code and OpenAI Codex
• Published Docker images cover linux/amd64 and linux/arm64 platforms
• Native binaries available for macOS (Apple Silicon and Intel) and Linux
• Windows support provided via WSL2 integration

This research explores how looped language models handle agentic tool use, specifically in compositional settings requiring coordinated API calls and state management. The study evaluates native and retrofitted models on benchmarks like API-Bank and BFCL, comparing them against non-looped counterparts under matched training conditions. Results indicate that recurrent computation significantly improves performance on tasks with dependencies, though gains are smaller for isolated API calls.

• Looped architectures excel at maintaining state and dependencies across multiple tool interactions.
• Recurrent computation provides measurable benefits for complex, compositional agentic workflows.
• Isolated API calls see minimal performance gains from looping, suggesting targeted deployment.
• Retrofitted looped models show promise, allowing upgrades without full retraining.

Standard prefix caching fails for agentic models because they repeatedly encode tool schemas in varying combinations. ReCache solves this by isolating resource representations into composition-invariant KV blocks via resource-wise attention. The framework further reduces overhead by pruning non-critical fields and restricting visibility to specific layer-head routes.

• Decouples tool schema encoding from request context to enable cross-request KV cache hits.
• Uses resource-wise attention to create composition-invariant KV blocks for better reuse.
• Reduces memory and compute via structural and semantic pruning of invocation-critical fields.
• Specific layer-KV-head routes limit visibility to only relevant computational paths.

This survey reframes LLM-based agents as systems with structural, dynamic states that evolve through memory, tools, and coordination. It argues that existing research treats graphs merely as support structures rather than as evolving substrates. The authors highlight that the coupling between evolving agent states and dynamic graph topology remains underexplored, particularly regarding how topology changes with feedback and new evidence.

• Shifts focus from static graph support to dynamic topology evolution in agent design
• Highlights gaps in understanding how agent state changes alter graph structure
• Suggests new perspective for building self-evolving multi-agent systems
• Connects memory and tool use directly to graph transformation mechanics

Researchers introduce FraudBench, a benchmark designed to evaluate conversational AI agents in banking contexts where they can execute financial actions like transferring funds or resetting PINs. Unlike existing benchmarks that focus on static transactions or generic prompt injection, this framework specifically tests how agents handle adaptive fraud attempts involving identity manipulation and trust exploitation over multi-turn dialogues. The study highlights the critical gap in current safety evaluations for agents that combine customer service with authorization capabilities.

• Existing benchmarks fail to test adaptive conversational fraud in policy-grounded agents.
• FraudBench targets identity manipulation and trust exploitation in banking dialogues.
• Agents with financial execution capabilities require specialized safety testing.
• Static transaction analysis is insufficient for dynamic conversational risk.

This paper explores using multimodal large language models for remaining useful life estimation in predictive maintenance. The proposed framework retrieves historically similar degradation segments and converts them into visual comparison artifacts alongside the test trajectory. These artifacts are processed by the MLLM via structured prompts to ground its predictions in concrete historical data patterns.

• MLLMs can be grounded for prognostics by retrieving similar historical time-series segments.
• Visual comparison artifacts bridge raw sensor data and multimodal model reasoning.
• Structured prompts guide the MLLM to analyze degradation trajectories effectively.
• Evaluated on C-MAPSS FD001, showing potential for agentic AI in maintenance.