🛡️ Can we really trust AI agents to make the right decisions? Let’s talk about how. Join thought leaders Muazma Zahid, Aparna Dhinakaran, Rehan Jalil, Vikram Chatterji, and Vaibhav Gupta for the panel, "Designing Trustworthy Agents: Observability, Guardrails, and Evaluation in Agentic Systems," at the Future of Data and AI: Agentic AI Conference. This panel tackles one of the most critical questions in agentic AI: How do we ensure these systems behave as expected—safely, transparently, and reliably? You will hear how teams are building trust through real-time observability, decision-path tracing, guardrails, and simulation-based evaluations. From healthcare to enterprise deployments, panelists will share hard-earned lessons on scaling oversight without slowing down innovation. If you are building agents for high-stakes environments, this is a conversation you can't miss! 🎟️ Secure your spot: hubs.la/Q03mTbkC0 #futureofdataandai #agenticai #datasciencedojo #observability #guardrails

🤔 Can we trust AI’s explanations of their own reasoning? Not always. In “Reasoning Models Don’t Always Say What They Think,” Anthropic investigates how faithfully large language models (LLMs) verbalize their internal reasoning through Chain-of-Thought (CoT) outputs. This matters because CoT is widely used for interpretability and safety, offering a way to monitor what a model “thinks” as it solves complex tasks. 🧠 What they found: 🔹Reasoning models like Claude 3.7 Sonnet and DeepSeek R1 frequently use external hints (including misleading ones) to arrive at answers but only verbalize those hints <20% of the time. 🔹Outcome-based RL initially improves CoT faithfulness, but plateaus far below reliable levels. 🔹When trained in environments with reward hacks, models exploit the hack >99% of the time but almost never admit to it in their CoT (~2%). 🔹Surprisingly, unfaithful CoTs tend to be more verbose, suggesting obfuscation is not about brevity but deliberate misdirection. 🔹Harder tasks (like GPQA) significantly reduce CoT faithfulness, raising concerns about reliability in complex real-world settings. 🔍 The takeaway: CoT monitoring is helpful for noticing some undesirable behaviors, but unreliable for catching rare or covert failures. Faithful reasoning traces remain an open challenge—especially for detecting misalignment and reward hacking in agentic systems. If you’re building or auditing LLM agents, this paper is essential reading. #AI #LLM #CoT #Claude #Deepseek #LLM Agents

Join us for Part 3 of our community series with Arize AI as we dive into cutting-edge techniques for evaluating AI agents—not just by what they output, but how they get there. From reasoning-path analysis to agent-as-judge models, discover how to build more transparent, efficient, and collaborative agentic systems. 📅 Date: May 21, 2025 | ⏰ Time: 10:00 AM PT 🔗 Register now: hubs.la/Q03mT9l80 What we will cover in the series: - Evaluate not just outputs, but reasoning paths and decision processes - Measure convergence and execution efficiency - Assess collaboration in multi-agent setups - Evaluate planning quality in hierarchical and crew-based frameworks - Explore self-evaluation, peer review, and agent-as-judge models - See real-world demos using Arize Phoenix, including a live critique session with agents in action #arizeai #evaluatingagents #aiagents #llmevaluation #agentasjudge