The State of Agentic Systems – September 2025
August proved agents could break out of chat boxes and start living inside apps and voice. September pushed further. Agents began transacting, running for hours at a time, and embedding directly into the systems enterprises already use.
OpenAI added real payment rails. Anthropic showed agents can sustain long sessions of work. Microsoft turned on Agent Mode in Office and gave enterprises model choice. Databricks partnered with OpenAI to put agents directly into data infrastructure. Even the infrastructure and policy news from NVIDIA, CoreWeave, and the GSA showed the ecosystem preparing for agents not as side projects but as core workflows.
The story of September is not about novelty or cleverness. It is about agents entering the places where money flows, compliance matters, and enterprises demand measurable results.
🔥 Spotlight Releases
OpenAI: Checkout and ACP
OpenAI introduced Instant Checkout in ChatGPT and released the Agentic Commerce Protocol with Stripe. This makes it possible for agents to complete real purchases. Etsy is live today, with Shopify merchants to follow. Importantly, ACP keeps merchants as the record holder so compliance and control stay intact.
Anthropic: Endurance and SDK
Anthropic launched Claude Sonnet 4.5, a model that can work for more than thirty hours at a time and posted strong scores on computer use benchmarks. Alongside it came an Agent SDK, the same backbone Anthropic relies on internally, now available for developers to build on.
Microsoft: Agent Mode in Office
Microsoft added Agent Mode to Excel, Word, and PowerPoint, while also introducing model choice. Enterprises can now route tasks to OpenAI, Anthropic, or Google models, making agents part of the tools workers already use every day.
OpenAI: Sora 2
OpenAI’s Sora 2 added a new app, feed controls, and moderation. It may look like a video tool, but the bigger picture is that video generation and editing are now programmable elements that agents can call.
Databricks and OpenAI
Databricks signed a one hundred million dollar plus partnership with OpenAI to integrate models directly into its platform. This brings agent capabilities into enterprise data workflows and simplifies procurement for large customers.
🛠️ Other Notables
AWS is turning on billing for Bedrock AgentCore, moving agent orchestration from pilot mode into a paid enterprise service. (Link)
Cartesia secured GDPR compliance for its voice platform, clearing the regulatory hurdle that often determines whether European companies can deploy voice agents at scale. (Link)
Vercel expanded access to its repo aware code review agent, which runs suggested patches in sandboxes before presenting them, pointing toward a next generation CI layer powered by agents. (Link)
Cisco wove conversational assistants directly into Webex, showing how incumbents will embed agents into existing collaboration flows instead of asking users to adopt new apps. (Link)
IBM previewed network assurance that uses agentic automation to detect and fix problems in real time, hinting at how critical infrastructure providers may trust agents for resilience. (Link)
Infrastructure and policy also made headlines. NVIDIA’s new multi-gigawatt deal with OpenAI and CoreWeave’s expanded $6.5B agreement show that compute providers are scaling capacity for agent workloads. Meanwhile, the U.S. General Services Administration issued calls for AI procurement guidance, signaling that agents are beginning to enter regulated government workflows.
🔍 What People Are Talking About
ACP and payment rails
Stripe and OpenAI published full specifications for the Agentic Commerce Protocol. These explain how payments and tokens are handled and how merchants remain the record holders. The bigger point is that ACP sets a standard for how agents can move money safely, making transactions auditable and trusted across platforms.MCP UI in practice
WorkOS released clear explanations of the Model Context Protocol UI. They showed how agents can return interactive components like product selectors or toggles and how these can pass information back to the agent securely. This demonstrates that MCP UI is maturing into a real framework for interactive workflows rather than a demo feature.'Memory benchmarks
LongMemEval and newer follow ups such as SGMem are becoming the common benchmarks for testing agent memory over long sessions. These standards matter because the ability to hold context for hours or even days is now critical for production use, not just lab experiments. Teams building durable agents are starting to be measured against these benchmarks.
🧠 September’s Big Theme
September was the month agents became transactional and durable. OpenAI enabled them to move money safely. Anthropic proved they can sustain work for days at a time. Microsoft brought them into Office, where business happens. Databricks and others made them easier to adopt and buy. The shift is clear. It is no longer about whether agents can think. The question now is whether they can transact, endure, and get adopted at enterprise scale.
September also marked the turn from interesting interfaces to revenue generating agents. Value is concentrating in three areas. Rails means the systems that allow safe and auditable transactions. Reach means embedding agents into the platforms enterprises already rely on, like Office or Databricks. Receipts means proving ROI with metrics such as cost per task, completion rate, and how often humans still need to step in.
As model performance levels out, advantage shifts to governance, integration depth, and the ability to handle checkout. The real scoreboard is no longer benchmarks. It is GMV per seat, cost per task, exception rates, and customer renewals. The bull case is agents becoming a new form of middleware that collects a small take rate on transactions, similar to payments companies. The bear case is runaway costs and compliance problems. Winners will be the teams that ship safe autonomy with clear limits, capture workflow data as proprietary feedback, and turn early pilots into ongoing paid contracts.
The next phase is not about agents thinking better. It is about agents closing sales, clearing tickets, and doing work that enterprises can trust and pay for.
🧛 Launching something agentic in October ?
Shoot me a note at priyanka@work-bench.com
Priyanka 🌊
I’m a Principal at Work-Bench, a Seed stage enterprise-focused VC fund based in New York City. Our sweet spot for investment at Seed correlates with building out a startup’s early go-to-market motions. In the cloud-native infrastructure and developer tool ecosystem, we’ve invested in companies like Cockroach Labs, Run.house, Prequel.dev, Autokitteh and others.