Hi all,

🎉 We have our first in-person event next week! Come by to talk about agents and meet the rest of the group.

Now that we're starting to wrap up 2025, the "year of AI agents," a few trends have emerged from the noise:

• Putting agents into production is harder than it seems
• Agents still have a long way to go on benchmarks
• Multi-agent systems are not always better than one agent
• Code agents (e.g. Claude) are emerging as generic agent platforms
• Research and production interest is shifting to payments and environments

Anthropic has deliberately moved towards a generic agent platform with the release of Claude Sonnet 4.5, launching the Claude Agent SDK to power "many other types of agents."

Aaron Levie at Box with another good agent take - this space moves so fast that you have to rebuild the tech all the time.

Looks good to me. I also wrote some thoughts on defining agents.

Events

• Oct 7: AGT NYC IRL Meetup #1
• Oct 4: AI Agents Hackathon
• Oct 6: Evaluating and Iterating on AI Agents
• Oct 9: Building AI Agents for the Real World
• Oct 16: AI Agent World Tour
• Nov 6: AI Agent 2 Agent Summit

News

• Silicon Valley bets big on ‘environments’ to train AI agents
• HKUST Launches World’s Largest AI-Powered Educational Sandbox Game
• Powering AI commerce with the new Agent Payments Protocol (AP2)
• Cloudflare Introduces NET Dollar to Support a New Business Model for the AI-Driven Internet
• OpenAI takes on Google, Amazon with new agentic shopping system
• AI Agents Are Getting Ready to Handle Your Whole Financial Life
• Notion launches agents for data analysis and task automation
• Amazon launches AI agent to help sellers complete tasks and manage their businesses
• Hopper's AI agent can book flights and cancel trips without human help
• AI agent hypefest crashing up against cautious leaders, Gartner finds

I think it's reasonable that by the end of the year, we'll have real applications of payments for agents, a selection of environment tools for agents, and an agent application in nearly every tech company. The momentum right now feels unstoppable, even with mixed results in production.

Fundraising

• Assort Health ($76M) - agents for healthcare
• Tabs ($55M) - agents for finance teams
• Motion ($38M) - agents for SMBs
• Mimica ($26M) - agent process learning
• Paid ($21M) - payments for agents
• Scalekit ($5M) - auth for agents
• Actual AI ($3M) - agents for engineering managers
• Scorecard ($3M) - continuous agent evaluation

The outline of the agent-based business is starting to take shape, with infrastructure pieces like auth and payments getting heavily funded. Vertical-specific applications (healthcare, finance, etc.) are primed to replace or augment entire departments. Nothing appears to be slowing down yet.

Posts

• Real AI Agents and Real Work by Ethan Mollick
• Coding as the epicenter of AI progress and the path to general agents in Interconnects
• Designing agentic loops by Simon Willison
• One year of agentic AI: Six lessons from the people doing the work by McKinsey
• Why AI agent projects are stalling in production in Computer Weekly
• Agent Auth: The Problem That Kills Production Agents by Arcade
• Why Multi-Agent Systems Fail by Galileo

Two significant milestones have been crossed recently: the ability of agents to actually deliver useful real-world results and the repositioning of code agents as general-purpose agents following lessons learned from coding usage. The pieces are being moved into place for an agent-first enterprise and we'll likely see the first big moves in that direction soon.

Projects

• Microsoft Agent Framework - agents for Python and .NET
• ROMA - Recursive Open Meta-Agents
• Blades - multimodal agents for Go
• AGNTCY - agent-to-agent collaboration
• Rig - agent framework in Rust
• Upsonic - safety-first high-performance agents
• Embabel - agent framework for JVM
• Cognee - memory for AI agents in 6 lines

The agent frameworks just keep on coming. Also, Agno has released the 2.0 version of its framework and Kestra has hit version 1.0. There is still apparently plenty of room for improvement in the space, with Microsoft recently joining the competition.

Learning

• Stanford CS329A: Self-Improving AI Agents
• Effective context engineering for AI agents
• Building an AI agent from scratch in Python
• Systems Engineering for Agentic Applications
• Agentic Design Patterns (pre-print preview here)

Some good learning material out there about agents has started coming out, which is a big improvement over the current approach (reading random blogs and GitHub repos). No canonical books or courses yet, but it is still early.

Papers

• Paper2Agent: Research Papers As AI Agents
• Agent²: An Agent-Generates-Agent Framework
• Virtual Agent Economies
• Psychologically Enhanced AI Agents
• Federation of Agents
• AgentGuard: Runtime Verification of AI Agents
• Framework for Studying AI Agent Behavior
• Training Agents Inside of Scalable World Models
• Overhearing LLM Agents
• MCP-AgentBench

Recent agent papers have been really exciting. It's starting to feel like the field is progressing beyond the low-level architecture of agents and towards using them as dynamic building blocks to solve all kinds of problems.

​If you know anybody else that would be interested in AGT NYC, please refer them to agtnyc.com to sign up.

Cheers,
Ivan