Thrilled to share a major milestone for @fiddler_ai! We’ve raised an $18.6M Series B extension, bringing our total Series B funding to $50M. This investment propels our mission to advance AI Observability and AI Safety, enabling enterprises to deploy trustworthy and scalable LLM.

RT @tobi: @levie Excellent topic. Memory portability and memory observability will be key. Memory is a graph. Your work related memory is….

RT @TheTuringPost: Small Language Models (SLMs) are the future of Agentic AI, claim @NVIDIA researchers. Moreover, they offer a method for….

RT @pritika_mehta: The truth about Soham Parekh. - the dude clears interviews, he’s good with that . - after clearing, he has junior people….

RT @Analyticsindiam: The AI revolution is being driven by leaders who are turning vision into real-world impact. Presenting the Top 20 CEOs….

We’re about to see AI systems that are too complex to trust without oversight. That oversight must be built in, not bolted on. The Control Plane for Compound AI is coming. And we’ll look back and wonder how we ever shipped agents without it. /7.

Clouds like @awscloud , @googlecloud, and @Azure provide the engines. Frameworks like @LangChainAI , @crewAIInc , and @DSPyOSS offer the orchestration and choreography. But we still need the runtime integrity, governance, and safety. That’s where @fiddler_ai fits. /6.

It’s not enough to look at Agent outputs after the fact. We need:. * Live trust evaluation (faithfulness, PII, hallucinations).* Runtime fallback logic.* Centralized audit trails.* CI/CD for agent workflows. That’s what makes AI enterprise-ready. /5.

When a system plans across steps, makes API calls, and updates databases… It stops being a “chatbot.” It becomes software with agency. Now ask yourself:. * Can we trace what it did?.* Can we stop it if it misbehaves?.* Can we prove it followed policy? /4.

Think of the following Compound AI Stack: . 🧱 The app = your LangGraph or CrewAI agents.🧠 The model = Claude, GPT-4, or Titan.🔌 The tools = @Zendesk , @Snowflake , @ServiceNow .🧭 The control plane = ???. Without a control layer, you’re flying blind. /3.

Compound AI ≠ bigger LLMs. It is systems composed of:. * Multiple agents.* Long-term memory.* Tool and API calls.* Real-world action. Like any distributed system, it introduces complexity, failure modes, and risk. /2.

We’re entering the next phase of AI adoption:. Not just prompts → responses. But compound AI systems: agents + ML + Gen AI that plan, retrieve, execute, and adapt. These systems are powerful but they’re also fragile. They need something we haven’t built yet: A Control Plane.

RT @fiddler_ai: At Scale AI & Data Conference on June 25, Krishna Gade from Fiddler AI will introduce Agentic Observability, a new approach….

What Makes Agentic Observability Different? Here are three beliefs guiding our work at @fiddler_ai :. 🔥 Hot Take #1: Logs and metrics in the traditional sense are dead. Agents require semantic tracing: understanding intent, memory, and beliefs in real time. 🔥 Hot Take #2:.

Here’s the core shift we're seeing:. Agents don’t fail like programs do. They fail like people do due to misalignment, ambiguity, or lack of oversight. Observability must move from surface-level telemetry to real-time introspection. /6.

For example:. A travel agent LLM plans a trip from NYC to Paris. It reflects thrice before locking dates. It miscalls a hotel API due to malformed params. Reroutes. Reflects again. Hands off to a booking agent. Everything was successful. But without semantic traces, you’d only.

At @fiddler_ai we call this Agentic Observability. It’s a new runtime stack that captures:. 🧠 Internal reflections.📋 Planning traces.🛠️ Tool-usage .🤝 Multi-agent coordination. It’s observability at the level of cognition, not infrastructure. /4.

Here’s what actually matters now when it comes to Agentic Observability:. a) Was the agent’s plan coherent across steps?.b) Did it mis-sequence tool calls?.c) Did it parse tool responses correctly?.d) Did it hand off tasks cleanly to sub-agents?.e) How many times did it reflect.

Agentic systems don’t just execute code, they reason. They generate plans dynamically. They revise their own thinking mid-task. They call APIs, chain tools, and maintain evolving belief states. A 200ms latency spike is the least of our worries here. /2.

Most observability tools were designed for deterministic applications. But the frontier is shifting fast. Agents now plan, reflect, use tools, and collaborate to solve open-ended tasks. These are not models. They are systems. And they break everything we know about monitoring.

RT @fiddler_ai: Which guardrails solution is right for your organization? One size never fits all — and the stakes couldn't be higher. Ou….