Agentic AI

This paper addresses latency waste in production LLM agents caused by regenerating procedural code on every request. It introduces an agentic tool-making pipeline that compiles repeated standard operating procedure steps into validated, versioned tools before deployment. By grounding synthesis in live execution traces and backend schemas, the system generates and repairs candidate tools, allowing runtime agents to call them directly instead of relying on inference-time coding.

• Replaces inference-time code regeneration with pre-compiled, versioned tools for repeated tasks.
• Tool synthesis uses live execution traces and backend schemas to ensure accuracy.
• Agents fallback to code generation only when pre-built tools are insufficient.
• Validated in a Fulfillment Center alarm-triage system with a 44-node SOP.

The AWS MCP Server now supports direct AI agent connections via AWS Sign-In and standard OAuth protocols, eliminating the need for custom authentication middleware. This update allows agents to leverage existing IAM permissions and governance controls while supporting both interactive browser-based and programmatic headless authorization flows. Administrators gain enhanced control through new IAM condition keys, token introspection APIs, and CloudTrail audit logging.

• Agents connect via standard OAuth without extra auth software
• Leverages existing IAM policies and AWS identities
• Supports interactive and headless authorization modes
• New IAM condition keys and token revocation APIs
• Full CloudTrail audit coverage for OAuth access

Researchers introduce DeepSearch-Evolve, a self-distillation framework designed to train tool-use agents that improve through their own experience. The system relies on DeepSearch-World, a deterministic environment featuring reproducible search and page-reading tools alongside 420K multi-hop QA tasks. This setup addresses the limitations of fixed teacher-distilled trajectories and sparse-reward reinforcement learning by enabling progress verification, grounded reflection, and failure recovery.

• DeepSearch-World provides a deterministic, verifiable sandbox for training web agents with reproducible tools.
• The dataset includes 420K multi-hop QA tasks generated via entity-level random walks.
• Framework supports self-evolution behaviors like progress verification and failure recovery.
• Bypasses weak supervision from sparse rewards by using iterative trajectory generation.
• Addresses long-horizon interaction challenges in agentic AI training pipelines.

This paper argues that aggregate safety metrics for planner-executor LLM pipelines conflate harmful intent reframing, planner refusal, and executor delegation. The authors introduce a five-condition controlled contrast design to isolate these mechanisms. Evaluated on synthetic and benchmark scenarios, the study reveals that overall pipeline safety is not a stable architectural property.

• Aggregate 'pipeline effect' metrics mask distinct failure modes in multi-agent systems.
• New five-condition design separates reframing, refusal, and delegation factors.
• Results show pipeline safety is unstable, not a fixed architectural trait.
• LLM-judged compliance used for evaluation across synthetic and benchmark sets.

This research demonstrates that biased LLM judges can effectively turn off the skill retirement mechanism in self-evolving agents. While retirement relies on accurate failure signals to prune poor skills, reference-free tasks force the use of judges that introduce systematic bias rather than random noise. The study proves this bias prevents the agent from correctly identifying and discarding bad skills, leading to uncontrolled drift in the skill library.

• Biased judges do not just add noise; they structurally disable the skill retirement mechanism.
• Reference-free task evaluation often forces the use of biased LLM judges.
• Skill drift occurs because agents fail to recognize and prune failing skills.
• Symmetric noise preserves retirement, but systematic bias breaks it.