Agentic AI

AWS Lambda durable functions has added native support for Pydantic AI, enabling automatic checkpointing of agent progress. This integration ensures that agents resume from the last completed step after interruptions like timeouts, rather than restarting entirely. Each model and tool call is treated as a durable execution step, preventing redundant processing. This approach significantly reduces costs for expensive, multi-step AI workflows by avoiding repeated execution.

• Pydantic AI agents automatically checkpoint progress to Lambda durable functions.
• Execution resumes from the last step after timeouts or failures, not from the start.
• Every model and tool call is treated as a durable step for precise recovery.
• Eliminates the need to manually write checkpoint and retry logic in code.
• Prevents costly redundant API calls in long-running document or research chains.

SearchAtlas transforms unstructured LLM search trajectories into structured graphs where edges track evidence propagation from initial queries to final answers. The framework uses an automated parsing pipeline that achieves an 86.0% edge F1 score against human annotations and maintains consistency across runs. By shifting focus from final-answer accuracy to the reasoning process, the tool reveals how credible evidence is retrieved to satisfy question constraints.

• Moves evaluation beyond final answers to analyze the actual search process.
• Graph edges explicitly model how evidence propagates through the reasoning trace.
• Automated parsing achieves 86.0% edge F1 against human-annotated ground truth.
• Reveals systematic differences in how agents retrieve and use evidence.

Auto-RecSys is an autonomous research agent designed to handle the specific challenges of training large recommendation models at scale. It addresses long feedback loops by enabling parallel exploration across multiple research directions, avoiding the bottlenecks of serial iteration. The system also manages high system complexity and fragile infrastructure dependencies through robust, recoverable execution of multi-day GPU jobs.

• Enables parallel hypothesis testing to overcome multi-day training feedback loops.
• Provides robust, recoverable execution for fragile infrastructure and complex configs.
• Autonomously manages long-horizon experimentation without constant human intervention.
• Reduces time-to-insight by automating iterative refinement of large models.

RobustSGPO extends semantic-gradient prompt optimization by explicitly defining edit scope, constructing patches, and verifying them before applying updates. The method allows the search to continue from either the current best state or retained snapshots, addressing previous gaps in edit operation resolution. Evaluations on the AgentX workflow show that periodic permission scheduling outperforms fixed limits, raising task completion from 60% to 80%.

• Specifies edit scope and validates patches to prevent unstable agent harness updates.
• Allows search continuation from incumbents or snapshots for better convergence.
• Periodic permission scheduling improves test scores over fixed maximum permissions.
• Task completion rates jump from 60% to 80% on held-out benchmarks.

This research introduces a systematic black-box framework for evaluating risks in autonomous agentic systems that execute multi-step tool calls. The approach utilizes a seven-domain taxonomy to map observable behaviors to specific risk categories and automates the generation of 120 adversarial scenarios per domain via SAGE-RT. Empirical validation across CrewAI and AutoGen architectures with four base models demonstrates the method's capability to uncover vulnerabilities that single-turn evaluations miss.

• Standard single-turn evaluations fail to capture multi-step agent vulnerabilities in production.
• The framework requires only basic system descriptions, enabling black-box risk assessment.
• SAGE-RT automates adversarial scenario generation across seven distinct risk domains.
• Human-validated LLM judges are used to assess the effectiveness of red teaming attacks.