Agentic AI

This paper investigates whether agentic AI architectures are superior to traditional protocols like REST, SOAP, and MCP for automating business tasks. The authors compare seven methods on a controlled arithmetic benchmark to evaluate trade-offs in development speed, performance, and cost. The study aims to clarify when the complexity of agents is justified compared to established alternatives.

• Compares agentic AI against REST, SOAP, and MCP on a 1,000-expression arithmetic benchmark
• Evaluates trade-offs between development speed, runtime performance, and operational cost
• Challenges the assumption that agents are universally preferable for automation tasks
• Provides empirical data on when simpler architectures may outperform agentic systems

Researchers demonstrate that LLM agent susceptibility to prompt injections is not fully determined at the moment malicious content is ingested. Using branching rollouts to replay episodes and resample subsequent steps, they isolate variance attributable to post-read reasoning. While non-thinking agents are largely compromised upon reading, enabling chain-of-thought reasoning shifts significant outcome variance to later decision steps, challenging the efficacy of static read-time monitors.

• Prompt injection outcomes are not fixed at ingestion; post-read reasoning significantly influences agent behavior
• Enabling chain-of-thought reasoning increases the share of outcome variance determined after reading malicious content
• Static runtime monitors at ingestion time are insufficient for agents using extended reasoning modes
• Branching rollout methodology effectively separates read-time vulnerability from reasoning-time susceptibility

A Meta engineer describes the growing unease among developers who ship AI-generated code without fully comprehending its internal logic. The rapid pace of AI-assisted development makes line-by-line verification impractical, forcing engineers to rely on high-level understanding rather than detailed code mastery. This shift conflicts with traditional engineering instincts that prioritize deep system knowledge.

• Line-by-line review of AI code is becoming time-prohibitive at scale
• Engineers are accepting high-level understanding over detailed code mastery
• Traditional instincts about code ownership clash with AI-assisted workflows
• Black-box AI outputs are normalizing despite reduced transparency

Researchers introduce Software World, a framework simulating multi-agent collaboration within realistic Python ecosystems. The system generates environments of interdependent packages where agents autonomously discover tasks and coordinate to improve software efficiency. This approach tests how agent teams scale inference-time compute through cooperative problem-solving.

• Simulates realistic multi-agent workflows using interdependent Python package environments
• Agents autonomously identify tasks and collaborate without predefined coordination protocols
• Framework evaluates collective agent performance on software efficiency goals
• Explores scaling inference-time compute through agent team structures

Researchers introduce Cortex, a supervisory layer for coding agents that enforces pre-execution governance and iterative repair. The paper formalizes requirement closure using finite lattice theory, proving that exhaustive iteration converges to a least fixed point under sound, monotone validation rules. It also provides stochastic bounds for completion time, though these guarantees rely on strict assumptions regarding validation fallibility and execution limits.

• Formalizes coding agent governance via finite lattice theory for requirement closure
• Proves convergence to least fixed point with bounded strict increases per requirement
• Offers stochastic bounds for completion time under uniform positive-progress assumptions
• Highlights critical dependencies on validation soundness and execution constraints