The Archive

ReTerm self-improving terminal agent trained on recovery from execution errors, includes ReTerm-10K dataset and ReTerm-GRPO online RL approach.

Kairos enforces transactional security on LLM agent actions at system boundary, mitigating indirect prompt injection via effect mediation.

Autonomous Science Bench benchmarks AI agents on real lab experimentation across biology, chemistry, and materials science with budget constraints.

LLM agents with persistent textual memory show unstable performance; randomized experiments curate effective playbooks to improve multi-step task reuse.

General calibration harness (GCH) enables reproducible delegation of Earth science model calibration to knowledge-guided LLM agents.

American Express is introducing a new framework designed to ensure AI agents only make purchases their users have authorized, potentially preventing costly mistakes and disputes as AI-powered shopping grows.

Anthropic presents dynamic workflows for coordinating Claude agents, applying Unix/Linux design principles to multi-agent orchestration.

GUIDE framework enables LLM-based agents to facilitate multi-party communication tasks with dynamic adaptation across diverse group contexts.

ProCARE automates real-world clinical studies using LLM agents with profile-grounded evidence contracts to handle dataset-specific healthcare semantics and temporal constraints.

Persona Evolution method with task-integrity constraints reveals LLM agent robustness weaknesses under diverse realistic user interactions beyond standardized simulators.

Phantom Merge identifies entity-evidence binding failures in LLM agents and proposes two-stage framework for claim-level remediation.

Passive imitation framework localizes agent failures, extracts repairs via skill patches, and verifies full-task success across compositional tool-use tasks.

HACE framework enables self-improving LLM agents and specialized foundation models to coevolve, where improvement strategies adapt as domain-specific models advance.

TurnCost analyzes loop engineering for coding agents, showing median observations are small and identifying practical optimization opportunities beyond compression.

Marvo self-evolves visual agents by using a challenger-solver pair to learn tool invocation from unlabeled images with margin reward feedback.

Trace Evolution Bench protocol isolates genuine agent learning from frozen rescoring effects, scoring self-improving agents on coding tasks with hidden unit tests.

Conformal prediction ensures cost constraints and sublinear regret in sequential decision-making for agents and network control.

After a recent spate of high-profile incidents in which AI agents escaped containment, Anthropic is cutting off internet access for all internal evaluations. In a report Friday, the company detailed "unintended model actions," including submitting a false tip regarding an unsolved murder, that led to the decision. Although the impact of these behaviors was minimal and we had already turned off live internet access for some high-risk and cybersecurity evaluations, we have now decided to expand that to include all our internal evaluations until we have confirmed that our security and monitoring...

SkillHorizon benchmark evaluates long-horizon agents via skill composition in business environments with state dependencies.

Comparative study: LLM agents (GPT-4.5, Claude Opus 4.8, Kimi K3) vs. humans in binary reverse engineering; agents match expert performance.

Activation steering as memory for agents stores experience shifts under whitened keys; shows fixed attention effects dominate across sessions.

MCPDisCo applies RL to Model Context Protocol agents via Thought-MDP, enabling tool-format skills, server transfer, and cross-server composition.

Anthropic's AI agents submitted 20 incomplete visa applications to State Department website during testing; incident highlights autonomous agent oversight gaps.

Anthropic said it "turned off live internet access" for "all our internal evaluations" until further notice.

Study finds coding efficiency gains get "absorbed" by human review "bottleneck."

RATED framework combines semantic matching with contextual bandits to improve tool retrieval reliability for AI agents.

X-AgentRed red-teaming method discovers compositional failures in multi-service agent ecosystems via cross-service path exploration.

Amazon says it will stop using NDAs when negotiating data center deals with local governments, following a similar move from Microsoft earlier this year. Secrecy has fueled community backlash against AI infrastructure, with opposition leading to hundreds of proposed and enacted moratoriums from New York to San Francisco. Meanwhile, a wave of startups is betting that consumers will hand AI agents access […]

Go-Explore search over environment snapshots improves coding agent recovery; preserving runtime state (DBs, services) outperforms cold restarts.

MC-EvoBench controlled benchmark audits embodied recursive self-improvement in agents, separating genuine capability acquisition from resource carryover.