Read the full blog post here:

🌍 Trusted by patent attorneys in 15+ countries, ClaimWise is moving faster than ever — spotting agent failure modes in 𝗱𝗮𝘆𝘀 𝗶𝗻𝘀𝘁𝗲𝗮𝗱 𝗼𝗳 𝘄𝗲𝗲𝗸𝘀 with Atla. “[𝘈𝘵𝘭𝘢’𝘴] 𝘳𝘦𝘢𝘭-𝘵𝘪𝘮𝘦 𝘥𝘢𝘵𝘢 𝘤𝘳𝘶𝘯𝘤𝘩𝘪𝘯𝘨 𝘢𝘯𝘥 𝘱𝘳𝘦𝘴𝘦𝘯𝘵𝘢𝘵𝘪𝘰𝘯.

👀.

RT @AgnoAgi: New integration alert!. @AgnoAgi 's production-ready multi-agent orchestration with the @Atla_AI sophisticated agent monitorin….

✒️ Full details in our blog post: .

🚀 𝗠𝗼𝗻𝗶𝘁𝗼𝗿 𝗮𝗻𝗱 𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗲 𝗮 𝗺𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄 — 𝗶𝗻 𝘁𝗵𝗶𝘀 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝘃𝗲 𝗱𝗲𝗺𝗼. Among teams using our platform that build with frameworks, one has stood out as the most popular: @AgnoAgi. We built a demo to show the combo:.→.

🌕 Start evaluating with Selene 1: 🌔 Selene 1 (70B): 🌖 Selene 1 quantized:

Big news: we’re open sourcing Selene 1!. We’ve been inspired by @huggingface's new native @vllm_project support for inference endpoints. We’re celebrating by making Selene our 70B LLM Judge model available to the wider community. Links 🧵.

Why do deep research agents fail? . We ran Open Deep Research through GAIA, a benchmark that requires web browsing and reasoning. The biggest issues: planning errors and reasoning errors. So, what can we do about it? We break it down in our latest blog post.

Researcher @sashankpisupati showed off a research demo at @AITinkerers earlier this week: a human-in-the-loop evaluator that helps stop agents from taking questionable actions. The task now is getting AI evaluators to do the same. Lots of shared pain around agents and their

RT @AAyman_1302: Got a room of 60+ people from @cohere @Synthesia @11x_official and more for the second @_ai_collective demo night in Londo….

Building with AI agents? The framework you choose will lay the foundation for how your agents think and act. We broke down 9 top open-source frameworks, from LangGraph to Google ADK, to help you find the right fit (link in comments). 🤔 Structured vs flexible .🤔 Lightweight vs.

We are working to automate this improvement layer, with promising early findings. Stay tuned and follow for more 🫡.

Most errors were terminal. Agents rarely recovered on their own. So how do we help agents recover?. We tested an actor-critic loop:. The agent acts. A critic evaluates each step. The agent self-corrects. With human critics, task success jumped 30%, reaching 80–90% accuracy. No

First, we diagnosed agent error types at the step-level. In DA-Code, the most common error category was reasoning errors (incorrect logic, hallucinated information, and not following instructions). Want to understand why your agents fail? Get in touch:

Before we can fix broken AI agents, we have to know how they’re breaking. We analyzed failures from DA-Code, a coding benchmark for data science tasks, to understand how agents break and what we can do about it. Here’s what we found 🧵.

AI agents are powerful, but still fail in unpredictable ways. Plus, debugging them is a mess. We just published a blog breaking down why agents still fail and what better observability could unlock. Check it out 🧵