AI Agents

This paper introduces AGENTSERVESIM, a simulator for evaluating LLM agent serving infrastructure. The authors argue that existing simulators are designed for stateless request-response workloads and fail to model the unique dynamics of agents: multi-turn interactions, gaps caused by tool execution, and opportunities for KV-cache reuse across turns. The new simulator models the entire agent program execution, including hardware awareness and memory hierarchies, allowing for cheap and fast evaluation of complex scheduling and caching policies. It reportedly matches real-system performance within 6% error.

The SecureClaw paper proposes a dual-boundary architecture to secure tool-using agents against two critical failure modes: unauthorized actions and internal data leakage. The system isolates the agent runtime by placing security controls at the I/O boundaries. Writes to external systems use a PREVIEW→COMMIT protocol where only a trusted executor can commit the action. Reads of sensitive data are passed through a gateway that replaces raw values with opaque handles, preventing the agent from ever directly seeing the secrets it's working with.

When an agent fails silently, how do you find the root cause in a long execution trace? The REFLECT paper proposes a novel method for error attribution that goes beyond passive observation. It works by actively intervening: a suspect step is identified, a patch is generated for the hypothesized error, and the trace is replayed with the patch. A successful outcome in the patched run provides strong contrastive evidence to pinpoint the original failure, improving localization accuracy, especially in complex tool-use scenarios.

The Anything2Skill framework proposes a method to move beyond simple retrieval-augmented generation (RAG). Instead of just retrieving fragmented declarative facts at runtime, it compiles procedural knowledge from heterogeneous sources (manuals, logs, examples) into a structured, reusable SkillBank. This offline process extracts, validates, and organizes skills into a persistent memory. At inference time, the agent can then retrieve not just relevant text passages but complete, executable skills, combining declarative knowledge with learned procedures.

This paper makes a strong case that uncertainty estimation for code generation cannot just reuse methods from natural language. It argues that code has unique properties—token fragility, the gap between intent and implementation, and executability—that demand a specialized approach. The authors propose a three-axis framework for measuring uncertainty: lexical (token entropy), algorithmic (consistency of generated pseudo-code), and functional (behavioral consistency across samples). An ensemble of these signals significantly outperforms standard NLP-derived uncertainty metrics for code.

Intuned is a new platform for building and running browser automations, pitched as a way to fill API gaps for data scraping or form submission. The core idea is that automations are generated by an AI agent but run as code, not as brittle UI scripts. A key feature is the "self-healing" capability, where the agent uses the context from runtime failures to debug and maintain the underlying automation code when a target website changes. This aims to provide the reliability of code without the manual maintenance burden.

Command Center pitches itself as an agentic coding environment for developers who care about quality. Its premise is that while AI can generate code quickly, developers now spend an inordinate amount of time reading and "de-slopping" it. The tool appears to provide an interface for managing multiple, large-scale, AI-generated diffs simultaneously. The core feature is a one-click "Refactor" button designed to turn the initial AI-generated output into higher-quality code, though the underlying mechanism for this is not detailed in the announcement.

Simon Willison offers a "believe it when I see it" take on Apple's 2026 AI announcements, colored by the over-promises of 2024. He notes the new Siri AI appears more feasible, leveraging a custom Gemini-derived model and vision capabilities to read the screen—a clever way to sidestep explicit app integration. While skeptical, he sees the new Core AI library and its PyTorch integration as a promising development for engineers wanting to run their own models on Apple hardware, a long-awaited capability.