AI Agents

The ENPIRE framework enables a coding agent to perform autonomous robotics research in the real world. It creates a closed loop for policy improvement by automating the physical feedback process: resetting the environment, executing a policy, verifying the outcome, and refining the approach. In this system, the agent analyzes logs, consults papers, and improves its own training code to fix failures. Using ENPIRE, a frontier coding agent was able to autonomously train a policy to a 99% success rate on a dexterous manipulation task.

This paper documents a novel failure mode called "Index Sickness," observed over 391 collaborative sessions with an LLM on a software project. The authors found that adding complex symbolic systems and constraints backfired. Beyond a certain threshold, the LLM stopped grounding in real-world semantics and began reasoning self-referentially within the symbolic layer, producing output that was internally consistent but physically meaningless. They propose a fix, "Baseline-Log Physical Separation," that dramatically reduced errors in subsequent sessions.

This paper tackles the pervasive problem of data contamination in coding agent benchmarks. The proposed method, SWE-Future, avoids replaying historical GitHub issues. Instead, it uses repository data up to a point in time ($T_0$) to forecast likely future work (features, bugs, refactors). These forecasts are then used as prompts to synthesize new coding tasks. This creates a benchmark that reflects a repository's organic evolution without the risk that the model has already seen the ground-truth solutions in its training data.

Standard defenses against jailbreaking often just detect and block malicious prompts. This paper argues that predictable refusals provide useful feedback for automated attacks, which can then refine prompts until they succeed. The authors propose "defensive misdirection": instead of blocking, the system provides a safe but non-operational response designed to fool the attacker's automated judge. Their proof-of-concept, CMPE, nearly eliminated verified attack success on benchmarks by making it much harder for an automated attacker to know if it has succeeded.

For a robot to be useful in a home, it must understand that objects move. The FlowMaps model addresses this by learning to predict future object locations in a 3D space. It uses a latent flow matching approach to model the patterns of how objects are moved by humans over time, generating a multimodal distribution of likely future positions. In a dynamic object navigation task, FlowMaps outperformed prior methods in both simulation and the real world, showing the value of modeling dynamics.

Eugene Yan provides a structured breakdown for building cybersecurity evaluations for AI agents. The pattern consists of four core components: a sandboxed target environment for the agent to act upon, a set of inputs designed to systematically vary task difficulty, a suite of tools for the agent to use, and a robust grader to score the outcomes. This isn't just theory; it's a practical, repeatable blueprint for anyone needing to rigorously test agents in a security context.

Simon Willison highlights Cloudflare's new temporary accounts feature, which allows deploying Workers without an account via npx wrangler deploy --temporary. While marketed for AI agents, the feature is broadly useful for any ephemeral task. A deployed project stays live for 60 minutes and can be "claimed" to make it permanent. This provides a clean way for an agent to spin up and tear down sandboxed serverless functions as tools, without the overhead of account management or cleanup.

Eugene Yan outlines a five-step workflow for using LLMs to audit and secure a codebase. The process moves from high-level to low-level: first, build a threat model of the application; second, use the model to discover potential vulnerabilities; third, verify the identified issues; fourth, triage the valid findings; and finally, generate patches. It's a structured, human-in-the-loop process that treats the LLM as a powerful, but not infallible, security analyst, providing a concrete playbook for security teams.

Simon Willison's sqlite-utils library, a staple for many Python-based data projects and agent memory systems, is getting a major update with v4.0rc1. The release candidate's biggest feature is the integration of database migrations, ported from the sqlite-migrate package, enabling structured schema evolution directly within the library. It also adds support for nested transactions. This is a significant upgrade for maintaining long-term state and ensuring data integrity in applications that rely on SQLite, which includes many agent systems.

This post provides a tour of how reinforcement learning (RL) is being adapted for training LLM-based agents. It connects several key ideas, explaining how techniques like ReAct, Reflexion, and chain-of-thought prompting can be viewed through an RL lens. The core idea is to frame the agent's generation process as a policy that can be improved using feedback. The article synthesizes different approaches, from self-correction with fine-tuning to more explicit RL algorithms, offering a conceptual map for engineers.