AI Agents

This paper delivers a crucial, and unsettling, result: agent safety hardening is domain-specific. The authors find that while frontier models are highly resistant to prompt injection in the web browser domain (0/140 successes), that resistance does not transfer to other contexts. The same models and attack families find success in a coding-agent domain. The work argues that headline attack success rates are misleading without understanding the specific domain and that model safety is not a monolithic, generalizable property.

AgentRedBench is a new benchmark addressing indirect prompt injection via SaaS tool integrations, a critical production threat. It uses an LLM to dynamically generate 215 subtle attack scenarios across 24 enterprise tools. Results are stark: even Claude Sonnet 4.6 fails 32% of the time without guards. The authors also release AGENTREDGUARD, a defensive model trained on adversarial tool-response data, which cuts the average attack success rate across models from 70% to just 2.4%, providing a practical defense.

This paper recasts agent memory consistency as a database concurrency control problem. It introduces TOKI, a bitemporal operator algebra that provides a formal contract for the ad-hoc heuristics (like last-writer-wins) used in production systems. By modeling belief updates with explicit isolation levels and provenance, TOKI formalizes how to resolve contradictions without admitting write-time anomalies like audit erasure or belief drift, which the paper shows plague existing systems that keep an LLM judge on the write path.

HarnessFix offers a framework for automatically diagnosing and repairing the harness that surrounds an agent, not just the agent's own logic. It compiles execution traces and harness code into a unified intermediate representation (HTIR), allowing it to attribute failures to specific steps and layers of the system. This enables targeted, automated repairs, a significant step up from blind prompt-tuning. On benchmarks like SWE-Bench, HarnessFix shows concrete improvements by systematically fixing underlying framework flaws.

RealClawBench tackles the problem of synthetic benchmarks by creating reproducible, scored evaluation tasks directly from real developer sessions with the OpenClaw agent. By reconstructing environments and using deterministic scorers, the framework converts messy, real-world requests into a controlled benchmark suite. The initial release of 281 tasks shows even the best models solve only 65.8%, highlighting a substantial gap between current agent capabilities and the difficulty of real-world software engineering tasks.

This paper provides a quantitative framework for "tokenomics" in agentic software engineering. The authors analyze where tokens are consumed across different operational phases like planning, tool use, and code generation. For anyone running code agents in production, this offers a much-needed model for understanding and optimizing the single biggest driver of opex: token costs. It elevates the conversation from "it's expensive" to a structured analysis of precisely why and where it's expensive, enabling targeted optimization.

OpenAI introduces "harness engineering," a conceptual framework for building robust systems around code-generating models. The approach emphasizes creating a structured environment—the harness—that provides context, manages execution, and verifies outputs. It's an explicit acknowledgment that the model is just one component in a larger system responsible for reliable agentic behavior. This signals a necessary shift from prompt-centric thinking to system-level architecture, especially for production code agents.

In a detailed practitioner report, an engineer from Jane Street describes a significant workflow shift: using Claude for UI development more than a traditional tool like Figma. The post documents an iterative, conversational process for generating and refining UI components directly into code. This is a powerful example of agent-assisted development moving beyond simple bug fixes and into the core design and prototyping loop, effectively collapsing the design-to-code gap for certain tasks and teams.

This post offers a deep architectural guide contrasting two philosophies for multi-agent systems: orchestrated breadth versus compounding depth. Using the open-source OpenClaw and Hermes frameworks as concrete examples, it frames the design choice as a critical strategic decision for any enterprise building a complex agent stack. It's a high-level strategy piece that's grounded in specific, contemporary tools, arguing this is the key design pattern to watch for the coming year.

Sebastian Raschka has released his curated list of notable LLM research papers from the first half of 2026. While the post itself is a collection of links and brief summaries, its value lies in Raschka's well-regarded curation. For engineers trying to keep up with foundational research without getting overwhelmed by the volume of preprints on arXiv, this serves as a high-signal starting point. The real technical depth, of course, is in the linked papers themselves.