Agents are talked about everywhere. Agentic is the new 2025 buzzword. OpenAI is going heads on against Anthropic when it comes to convincing developers to build on top of their AI empires.
Press enter or click to view image in full size
OpenAI’s Agents Strategy
OpenAI started first with GPTs. Most people only looked at the upload capabilities of documents that ChatGPT could access to answer specific questions. The real power of GPTs were one step deeper and are called Actions. Actions allow OpenAPI integrations (v3) with oAuth and other security mechanisms. This enables developers to integrate any REST APIs and use them together with ChatGPT. You can search for tickets, flights,… Only supporting v3 and not v2 meant that more established APIs, e.g. Amadeus for flight bookings, we’re not working.
The next service OpenAI launched was Operator. Operator is best described as robotics process automation on steroids. If you can access it via a browser then a human plus Operator can form an amazing team. The human might have to get involved with Captchas, logins, payments,… but outside that Operator can do its thing. Operator is amazing to stitch together thousands of websites because it can interact like a human. Unfortunately given Operator uses the same graphical browser interface, speed at scale is an issue. An API integration would allow thousands of requests per second. To get to that level with Operator would become very expensive quite quickly. Also access to private Intranets or local files is not possible.
In comes the Responses API which enables developers to do web searches, upload local files, local Operator sessions. By running a local agent inside a company’s firewall, local files and intranets can be accessed as well.
The Agents SDK, a more advanced version of the Swarm SDK, allows multiple agents to collaborate and to get to Agentic AI. Third-party APIs can be integrated with the Agents SDK, allowing for very advanced Agentic solutions.
The Anthropic Agents Strategy
Anthropic is the company behind Claude. They launched the Claude Desktop, which enables anybody to install a chat desktop client that talks to Claude and upload files and other local documents. By itself, this would not be a lot better than accessing ChatGPT from the browser, so Anthropic launched the Context Model Protocol or MCP. MCP enables developers to integrate any tool [APIs, databases, local file system tools,…], resource [large amounts of documents, images, videos,…] and prompts [structured bot conversations] as well as optionally have the code use the LLM for completions [here is some data can you please do this or that with it].
One of the core parts of MCP is that it is a standard which has been fully open sourced. Block.com open sourced Goose which allows different LLMs, e.g. OpenAI, Claude but also locally ran Llama, to integrate with many MCPs. The next version of MCP will also introduce advanced oAuth capabilities, making access to internal resources and tools a lot more secure.
Most MCPs are written in a Python or NodeJS and often random code gets downloaded from the Internet and executed on the user’s desktop. This is definitely not what CISOs in mid to larger companies will approve. They often blocked ChatGPT in the first place to avoid employees uploading confidential information.
The next step
Lots of companies will get behind OpenAI and will launch amazing solutions. This is similar to the time AWS was the majority Cloud in town and these same companies were using CloudFormation and lots of other proprietary AWS solutions. They locked themselves so deep into AWS that when spending started to go up, they could not do anything but keep on paying Amazon.
OpenAI is losing billions a year, unlike AWS, which is Amazon’s profit engine. What happens when VC money runs out? Prices will go up very quickly. If you are locked in, then you can not get out. Also, OpenAI has a very spotty history when it comes to service availability. Does a company want to really have its AI go down when ChatGPT is down, again, and again?
Although it took more time at the end, Terraform/OpenTofu are now more dominant than CloudFormation. In the same spirit, MCP is more likely to become the future standard. The MCP concept is more flexible with on/off premise and multiple LLM engines. However, I don’t think the current format will win. I am personally looking into packaging MCP into Wasm containers and running a complete MCP infrastructure in WasmCloud as standalone or inside Kubernetes. This way enterprises have a solution they can host themselves with their own LLMs or use third-party hosted solutions. Security, scalability, monitoring,… and all these “boring enterprise features” can be guaranteed.
If your company is planning to go into using or launching agents and you need some help, why don’t we talk? Also, if you want to join forces and either help with the open source MCP solution or have some prototypes created for your customers, please reach out. Investors are always welcome as well :-)
Anybody else? A clap or a follow, are really appreciated.