Agentic AI

AWS has generally released the Context Ontology Accelerator, an open-source tool designed to build machine-readable business ontologies for AI agents. The system ingests structured and unstructured data to draft an ontology, which domain experts then review and approve before it is stored in a W3C-standard knowledge graph. Agents access this trusted context via a Model Context Protocol (MCP) server to ensure decisions are consistent and auditable.

• Drafts ontologies from data using AI, requiring human expert validation for accuracy.
• Stores approved models in a W3C-standard knowledge graph owned by the organization.
• Exposes context to AI agents via a Model Context Protocol (MCP) server.
• Aims to improve trust, explainability, and auditability in agent decision-making.

A new study characterizes AI coding agent workloads using 761 million LLM calls from 3.2 million GitHub Copilot users. The data shows sessions consist of sparse user turns followed by autonomous loops of LLM inference and tool execution. This structure results in high KV cache hit rates within turns but significantly lower rates across turn boundaries.

• Agentic coding differs from chatbots with interleaved LLM inference and tool execution.
• Sessions feature sparse user turns leading to autonomous agent loops.
• KV cache hit rates average 90% within turns but drop to 55% across boundaries.
• Workload structure has distinct implications for systems design and optimization.

Researchers identify a small set of MLP neurons that linearly separate common agentic LLM tool-use errors: invalid arguments, unnecessary calls, and missing calls. They introduce PRISMS, a closed-loop framework that leverages these failure-specific neurons for both sparse detection and activation steering. Evaluated across Qwen3, Llama, and Gemma models, the method effectively targets over-calling and missing tool use patterns.

• Three key tool failures are detectable via linearly separable neuron activations.
• PRISMS unifies detection and steering using a shared sparse neuron basis.
• L1-regularized detectors fit on contribution-critical MLP activations.
• Framework validated on Qwen3, Llama, and Gemma model families.
• Sparse approach offers precise monitoring without full-model overhead.

This paper details a production deployment of a shared memory system designed to capture tacit enterprise knowledge that falls outside public training data or formal docs. The platform automatically collects task-adjacent experience with contributor approval, curating it into reusable question-answer pairs. This approach integrates knowledge capture directly into the coding workflow to prevent repeated rediscovery of internal conventions and fixes.

• Integrates knowledge capture into the coding workflow rather than relying on manual agent recording.
• Captures tacit enterprise knowledge like internal DSLs, local conventions, and recent fixes.
• Curates raw experience into reusable question-answer memories for retrieval by coding agents.
• Requires contributor approval to ensure quality and appropriate gating of shared memories.

The paper introduces AgentMemBench, a unified benchmark designed to evaluate long-term memory management in conversational AI agents. It compares five distinct strategies—in-context windowing, external key-value stores, graph-based episodic memory, compression-based summarization, and web-augmented memory. The assessment utilizes three public datasets covering multi-session dialogue, document grounding, and persona chat, measuring metrics like recall, faithfulness, and memory footprint.

• Evaluates ICW, EKV, GEM, CBS, and WAM under identical conditions for fair comparison.
• Tests across LoCoMo, MultiDoc2Dial, and MSC datasets for diverse long-term scenarios.
• Measures Recall@k, MRR, nDCG@k, Answer F1, LLM-judge Faithfulness, and Memory Footprint.
• Addresses the bottleneck of finite context windows in multi-turn conversational agents.