Emergence AI

Welcome to Emergence, where the future of enterprise workflow automation begins. Listen to exciting words from our co-founders Satya Nitta, Sharad Sundararajan, and Ravi Kokku, Learn Capital's founder and investor Rob Hutter, our research scientist Ashish Jagmohan, and our Chief Design Officer Hélène Alonso as they share how we’re advancing the science and development of #AIagents. Follow us to discover how intelligent agents will unlock the full potential of #AI in enterprise systems.

Web automation is more than browser clicks — it’s about dynamic planning, implicit and explicit search, and adaptive workflows. Let’s explore how hierarchical dynamic planning and tree search are tackling complex tasks 👇 Web automation is a key agentic capability for Emergence AI. Many general enterprise tasks can be accomplished via web-interface manipulation. There is also a universe of legacy apps that may only be controllable via web interfaces. Additionally, API-using agents can be combined with web automation to yield powerful agentic flows. Our state-of-the-art web agent, Agent-E, uses hierarchical dynamic planning to convert complex high-level user queries and tasks to (often long!) sequences of low-level primitive actions. A critical component of planning is search; planning systems search for sequences of actions to attain a specified goal or maximize rewards. Agent-E employs implicit search: its low-level browser navigation agent selects and executes actions, verifies results, and provides feedback to the high-level planner, allowing for replanning upon failure. This process enables Agent-E to search in the space of actions and recover from failures. There are, of course, other approaches to search-for-planning. Other web automation agents experiment with explicit search methods, often based on tree search, where the agent evaluates a set of candidate actions at each stage. This evaluation can rely on per-step AI critics validating the result of the action and/or on the basis of "rolling out" the entire subsequent sequence of states and actions and evaluating the final result. In both cases, the goodness of evaluated actions and browser states is maintained in a tree structure, driving the search for the next candidate action. Here are some examples of explicit tree-based search: 1. Koh et al. Tree Search for Language Model Agents uses a best-first tree search, achieving significant improvements on the VisualWebArena benchmark. 2. Agent-Q uses Monte-Carlo Tree Search, showing significant improvements on the WebShop dataset. 3. Google's Project Mariner reports significant improvement in task completion rate on the Web-Voyager dataset. But there still remain tradeoffs — a tree-based search needs to maintain several browser states and tends to have high latency when used at runtime. Nonetheless, it offers exciting opportunities for innovation and optimization, with significant implications for agent design and performance. #AI #WebAutomation #AgentE #MCTS #AutonomousSystems

We're pleased to announce the acceptance of our paper to the AI Agent for Information Retrieval (Agent4IR) workshop at the 39th Annual AAAI Conference on Artificial Intelligence 🚀 We will present 'Better RAG using Relevant Information Gain,' co-authored by our Research Scientists and Engineers Marc Pickett, Jeremy H., Ayan Kumar Bhowmick, Raquib Ul Alam, and Aditya Vempaty. This work introduces a state-of-the-art principled retrieval algorithm, Dartboard, that implicitly promotes diversity of retrieved passages by optimizing relevant information gain. We look forward to exploring cutting-edge AI research and impactful enterprise solutions with fellow presenters and attendees! #MultiAgentOrchestration #EmergenceAI #AAAI2025 Association for the Advancement of Artificial Intelligence (AAAI)

Impactful conversations on agents and #GenAI in research at #COMSNETS2025! Our Senior Vice President (Innovations) and India Head, Prasenjit Dey, joined an inspiring panel of leaders from academia and industry at COMSNETS to discuss "Research and Education in the Time of GenAI," which explored: ✔️ GenAI's potential to fundamentally transform how scientists and engineers approach research, emerging as an indispensable tool. ✔️ The ability of GenAI to advance the research process, including how we review literature, infer unseen concept connections, and present hypotheses, accelerating the discovery cycle for various applications. ✔️ How agentic systems are inherently well-suited for such iterative loops of Observation -> Hypothesis generation -> Validation -> Refinement. Thanks to COMSNETS and all the attendees for making this event such a success! Follow our journey as we drive productivity gains and unlock new possibilities with AI for enterprises. #AIagents #MultiAgentOrchestration #ScientificDiscovery

Discover how AI agents can adapt to dynamic and challenging enterprise environments👇 In our recent Live Q&A, our Research Scientist and Manager, Aditya Vempaty, joined our VP of AI Agents, Vivek Haldar, and VP of Developer Relations and Community, Waqas Makhdum, to discuss the launch of our enterprise-grade multi-agent orchestrator and its capabilities. Diving into how our multi-agent orchestrator balances the need for adaptability, determinism, and reliability, Aditya Vempaty shares: ✔️ ⁠How our orchestrator flexibly adjusts to complex enterprise environments, accurately assigning AI agents to specific tasks. ✔️ How the orchestrator strikes a balance between adaptability and determinism, focusing on reliable output. ✔️ How self-improvement enables agents to add new skills and align with evolving enterprise systems for long-term success. Explore real-life use cases with our multi-agent orchestrator and drive productivity gains for your enterprise 🚀 #AIOrchestration #MultiAgentSystems #EmergenceAI

Delighted to share a subset of E-Web: a new Web benchmark that our team at Emergence AI designed to rigorously evaluate Web agents. E-Web aims to set a new standard for assessing AI capabilities in real-world enterprise settings. We are releasing many prompts in the open source, and a paper that describes the methodology in developing the benchmark. Benchmark Paper: https://lnkd.in/e5fgFs4R Benchmark Github: https://lnkd.in/eNsXwk_8 What Sets E-Web Apart? - Skill-centric Design: Core, transferable skills such as form filling, dropdown handling, and file operations that are common across applications. - Real-World Relevance: Enterprise tasks based on tools like Salesforce, JIRA, and LinkedIn. In the evolving landscape of AI, there is a clear need for benchmarks that reflect enterprise use cases. While E-Web is a step forward, we believe that there is much more to explore. We look forward to seeing contributions that extend and enrich this benchmark. Heads-up: Multiple Web agents do poorly on these prompts. We will share the evaluation results soon. #WebAgents #EnterpriseAI #Benchmarks #AgentE #Agents

Thanks to Agency for featuring Emergence AI in their 2024 AI Agent Market Map! We recently launched our enterprise-grade multi-agent orchestrator — an autonomous meta-agent that can plan, execute, verify, and iterate in real time. It combines human-like interaction and navigation with machine-level scalability, enabling businesses to orchestrate operations across web front ends, APIs, and both modern and legacy enterprise systems, unlocking valuable use cases. Keep an eye out to see how we advance multi-agent orchestration in 2025 🚀 #EnterpriseAutomation #AIOrchestration #WebAutomation #EmergenceAI

🔎 𝐄𝐦𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐑𝐞𝐟𝐥𝐞𝐜𝐭𝐢𝐨𝐧𝐬 𝐒𝐞𝐫𝐢𝐞𝐬 | #5 𝐇𝐨𝐰 𝐝𝐨𝐞𝐬 𝐨𝐧𝐞 𝐮𝐬𝐞 𝐦𝐮𝐥𝐭𝐢-𝐚𝐠𝐞𝐧𝐭 𝐨𝐫𝐜𝐡𝐞𝐬𝐭𝐫𝐚𝐭𝐢𝐨𝐧 𝐭𝐨 𝐛𝐮𝐢𝐥𝐝 𝐦𝐢𝐬𝐬𝐢𝐨𝐧-𝐜𝐫𝐢𝐭𝐢𝐜𝐚𝐥, 𝐫𝐞𝐥𝐢𝐚𝐛𝐥𝐞 𝐞𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐬𝐲𝐬𝐭𝐞𝐦𝐬? * 𝐄𝐱𝐩𝐥𝐨𝐫𝐚𝐭𝐢𝐨𝐧 𝐯𝐬 𝐄𝐱𝐩𝐥𝐨𝐢𝐭𝐚𝐭𝐢𝐨𝐧: Agentic systems with their inherent stochasticity can be very useful to explore new design spaces for multi-agent orchestration to transform enterprise workflows. Once the design is validated by target performance metrics or humans, it can be reliably exploited again and again in operations time. * 𝐑𝐞𝐜𝐨𝐯𝐞𝐫𝐲 𝐟𝐫𝐨𝐦 𝐞𝐫𝐫𝐨𝐫 𝐦𝐨𝐝𝐞𝐬: While a validated mult-agent orchestrator design can be exploited repeatedly for a deterministic and predictable operation, which is important for most enterprise workflows, it still does not fulfill the vision of a “reliable” agentic system. When an error mode is encountered, the orchestrator should be able to quickly re-plan and reconfigure the system to ensure that normal operations of the system can be quickly restored. * 𝐇𝐮𝐦𝐚𝐧-𝐢𝐧-𝐭𝐡𝐞-𝐥𝐨𝐨𝐩 𝐢𝐬 𝐭𝐡𝐞 𝐤𝐞𝐲: As agentic systems dynamically reconfigure themselves to recover from errors, it is extremely critical that there is a well-defined Role-based Access Control (RBAC) for agents (as much as humans) while deploying multi-agent orchestration systems in the enterprise. Based on the criticality of the change operation (Read, Write, Modify, Delete), the orchestrator should be able to approve them without human intervention or fall back on humans for approval when any change can: - Have a significant impact on the core process - Result in regulatory compliance issues - Cause integrity issues to the core data for the process - High financial impact * 𝐒𝐞𝐥𝐟-𝐢𝐦𝐩𝐫𝐨𝐯𝐞𝐦𝐞𝐧𝐭: As the system collects more and more approvals/rejections from humans during operations and recovery from errors/outages, it learns from these experiences and falls back lesser and lesser on humans over time. This ensures a graceful evolution of the system towards limited autonomy without compromising on reliability, which is so important in enterprise operations. #EmergenceAI #AgentsInEnterprise #SelfImprovingAgents

🔎 𝐄𝐦𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐑𝐞𝐟𝐥𝐞𝐜𝐭𝐢𝐨𝐧𝐬 𝐒𝐞𝐫𝐢𝐞𝐬 | #4 𝐖𝐡𝐲 𝐢𝐬 𝐩𝐥𝐚𝐧𝐧𝐢𝐧𝐠 𝐜𝐫𝐢𝐭𝐢𝐜𝐚𝐥 𝐢𝐧 𝐦𝐮𝐥𝐭𝐢-𝐚𝐠𝐞𝐧𝐭 𝐨𝐫𝐜𝐡𝐞𝐬𝐭𝐫𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐞𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐰𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬? Planning and executing complex workflows in enterprise settings such as #compliance, #QA, #research, and #ProjectManagement comes with unique challenges: balancing reliability, cost-efficiency, flexibility, and robustness to unexpected inputs and results. Our orchestrator platform addresses these challenges with innovative solutions: * 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐭 𝐏𝐥𝐚𝐧𝐧𝐢𝐧𝐠: Breaks down complex intents into discrete steps, dynamically generating and executing plan steps to ensure resilience against error and allow replanning where necessary. * 𝐑𝐞𝐮𝐬𝐚𝐛𝐥𝐞 𝐖𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬: Stores successfully executed plans for retrieval, boosting consistency, reducing latency and cost, and allowing for parallel execution of steps. * 𝐑𝐨𝐛𝐮𝐬𝐭 𝐕𝐞𝐫𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧: Ensures quality with final-output checks, step-level verification, and human-in-the-loop oversight when needed. Learn more about how we’re streamlining planning and multi-agent collaboration for smarter, faster, and more reliable outcomes for #MultiAgentOrchestration in enterprise workflows: https://lnkd.in/egr6NJae

🔎 𝐄𝐦𝐞𝐫𝐠𝐞𝐧𝐜𝐞 𝐑𝐞𝐟𝐥𝐞𝐜𝐭𝐢𝐨𝐧𝐬 𝐒𝐞𝐫𝐢𝐞𝐬 | #3 𝐇𝐨𝐰 𝐜𝐚𝐧 𝐰𝐞 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐞 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐝𝐞𝐬𝐢𝐠𝐧𝐞𝐝 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐰𝐞𝐛 𝐢𝐧 𝐞𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐞𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭𝐬 𝐦𝐨𝐫𝐞 𝐞𝐟𝐟𝐞𝐜𝐭𝐢𝐯𝐞𝐥𝐲? The web is a dynamic and multifaceted landscape, and when applied to enterprise scenarios, the challenges for AI agents become even more complex. With E-Web, we introduce a new benchmark for assessing how AI agents navigate and execute real-world web tasks that are tailored to meet the rigorous demands of enterprise applications. Dive into the details here: https://lnkd.in/e5RA7jPn #AIResearch #AIAgents #EmergenceAI