AI Agents — From Inspiration to Real-World Insights: What’s Working and What’s Not (Notes from GenAI Summit 2025)

After five intense days at GenAI Summit 2025, one theme stood out across the board: Agentic AI. From blue-sky visions to production challenges, the conversation around AI agents is evolving fast — and here's a summary of the key takeaways from my notebook:

🔮 The 2028 Workday Vision - Wordware

CEO Filip Kozera sees it as: AI takes on the routine, while humans bring strategy and judgment.

• Personal agents handle inbox triage, assign tasks to other agents, and offer leadership feedback based on meetings and books.
• Integration into tools like Slack boosts productivity — agents quietly approve tasks behind the scenes.

🧠 The Future of Workplace AI — Glean's Vision

With over 100 million agent actions annually, Glean’s CEO Arvind Jain shares a compelling vision:

• A deeply personal AI companion that observes your workflow and helps orchestrate your day.
• Key focuses: security, agent identity, guardrails, and intuitive interfaces (UI/voice).

🏗️ Building Enterprise Agents – Glean’s Journey

• Started in 2019, using transformer models for enterprise knowledge search. Solved the "60% time spent asking colleagues" problem.
• Early traction came from evangelists and tight integration with tools like Confluence.
• Now supports workflows like performance reviews by surfacing relevant artifacts.

Challenges? Integration, compliance, security, and preserving institutional memory. Evolution: From Q&A bots to action agents (e.g., approving invoices), while complementing — not replacing — documentation.

⚙️ Agentic AI Inference – Friendli

Agentic AI intensifies traditional AI infra demands: long sessions, tool orchestration, multi-modal inputs. Friendli AI offers high-performance model inference with low latency, high throughput, and GPU efficiency — critical for scaling agent capabilities.

🧠 Memory & Context Limits - Mirix

Agents need to track long-term context, but token limits remain a real constraint. Professor Xi Chen introduced Mirix's solution.

Mirix tackles this with the only memory system for true multi-modal context — visual + language.

• Outperforms RAG by 35%, while cutting storage by 99.9%.
• Categorizes memory into: Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault.

🧩 Big Brain Theory — VC Advice to Founders

• Build vertical agents to automate deep workflows (e.g., cutting sales overhead by 70%).
• Focus on trustworthy results, not just LLM scale.
• Use smart chunking strategies to maintain contextual integrity.

✅ Trust, Evaluation & Scaling

• Trust should be rooted in values, not just models. Evaluation tools must track bias, intent, and truth vs. engagement metrics.
• LLMs aren’t always the answer — especially for mission-critical tasks.
• Despite the hype, only ~3 million enterprise users actively use agents today. Adoption hinges on the user's skill, not model strength.

⚖️ Regulation, Performance & Reality Checks

• Autonomous agents need standards, evaluations, and openness to avoid LLM lock-in.
• In Similar AI's 200 steps tasks test, today’s LLM agents hit 41.4% success rates, far below humans at 72%.
• CEO Ang Li talked about Similar AI uses Neural-symbolic methods and humans-in-the-loop ensure determinism and trust.
• Local models running on laptops prove that you don’t need billion-dollar budgets to build useful agents.

🛠️ Real-World Agent Use Cases

Agents are already doing:

• Payroll management
• Web shopping
• Stripe chargeback resolution
• Arbitrage tasks

With the right oversight, these agents can free up time — while keeping humans in the loop to drive creativity and decision-making.

Final Thought: Agentic AI has evolved from a blue-sky concept to a deployed reality—it's now a frontier where bold ideas meet operational complexity. The road ahead demands stronger infrastructure, smarter memory systems, and thoughtful regulation. But the momentum is real.

#GenAI #AgenticAI #AIagents #EnterpriseAI #LLM #Inference #MemorySystems #FutureOfWork

♻️ Repost if this hit home.

🔔 Follow Nicole Hu for more AI insights.