Context Serialization

In a caller variation of The Sequence Engineering newsletter, “Why Did MCP Win?,” the authors constituent to discourse serialization and speech arsenic a reason—perhaps the astir important reason—why everyone’s talking astir the Model Context Protocol. I was puzzled by this—I’ve work a batch of method and semitechnical posts astir MCP and haven’t seen discourse serialization mentioned. There are tutorials, lists of disposable MCP servers, and overmuch much but thing that mentions discourse serialization itself. I was adjacent much puzzled aft speechmaking done the MCP specification, successful which the presumption “context serialization” and “context exchange” don’t appear.

What’s going on? The authors of the Sequence Engineering portion recovered the bigger picture, thing much important than conscionable utilizing MCP to fto Claude power Ableton. (Though that’s fun. Suno, beware!) It’s not conscionable astir letting connection models thrust accepted applications done a modular API. There isn’t a abstracted conception connected discourse serialization due to the fact that each of MCP is astir discourse serialization. That’s wherefore it’s called the Model Context Protocol. Yes, it provides ways for applications to archer models astir their capabilities truthful that agents tin usage those capabilities to implicit a task. But it besides gives models the means to stock the existent discourse with different applications that tin marque usage of it. For accepted applications similar GitHub, sharing discourse is meaningless. For the latest procreation of applications that usage networks of models, sharing discourse opens up caller possibilities.

Here’s a comparatively elemental example. You whitethorn beryllium utilizing AI to constitute a program. You adhd a caller feature, trial it, and it works. What happens next? From wrong your IDE, you tin telephone accepted applications similar Git to perpetrate the changes—not a large deal, and immoderate AI tools similar Aider tin already bash that. But you besides privation to nonstop a connection to your manager and squad members describing the project’s existent state. Your AI-enhanced IDE mightiness beryllium capable to make an email. But Gmail has its ain integrations with Gemini for penning email, and you’d similar to usage that. So your IDE tin bundle everything applicable astir your discourse and nonstop it to Gemini, with instructions to determine what’s important, make the message, and nonstop the connection via Gmail aft it has been created. That’s different: Instead of an AI utilizing a accepted application, present we person 2 AIs collaborating to implicit a task. There tin adjacent beryllium a speech betwixt the AIs astir what to accidental successful the message. (And you request to corroborate that the effect meets your expectations—vibe emailing to a brag seems similar an antipattern.)

Now we tin commencement talking astir networks of AIs moving together. Here’s an illustration that’s lone somewhat much complex. Imagine an AI exertion that helps farmers program what they volition plant. That exertion mightiness privation to use:

• An economics work to forecast harvest prices
• A work to forecast effect prices
• A work to forecast fertilizer prices
• A work to forecast substance prices
• A upwind service
• An agronomy exemplary that predicts what crops volition turn good astatine the farm’s location

The exertion would astir apt necessitate respective much services that I can’t imagine–is determination an entomology exemplary that tin forecast insect infestations? (Yes, there is.) AI tin already bash a bully occupation of predicting weather, and the fiscal manufacture is utilizing AI to bash economical modeling. One could ideate doing this each connected a giant, “know everything” LLM (maybe GPT-6 oregon 7). But 1 happening we’re learning is that smaller, specialized models often outperform ample generalist models successful their areas of specialization. An AI that models harvest prices should person entree to a batch of important information that isn’t public. So should models that forecast effect prices, fertilizer prices, and substance prices. All of these models are astir apt subscription-based services. It’s apt that a ample farming concern oregon cooperative would make proprietary in-house models.

The farmer’s AI needs to stitchery accusation from these specialized models by sending discourse to them: what the husbandman wants to know, of course, but besides the determination of the fields, upwind patterns implicit the past year, the farm’s accumulation implicit the past fewer years, the farm’s technological capabilities, the availability of resources similar water, and more. Furthermore, it’s not conscionable a substance of asking each of these models a question, getting the answers, and generating a result; a speech needs to hap betwixt the specializer AIs due to the fact that each reply volition power the others. It whitethorn beryllium imaginable to foretell the upwind without knowing astir economics, but you can’t bash cultivation economics if you don’t recognize the weather. This is wherever MCP’s worth truly lies. Building an exertion that asks models questions? That’s decidedly useful, but immoderate precocious schoolhouse pupil tin physique an app that sends a punctual to ChatGPT and screen-scrapes the results. Anthropic’s Computer Use API goes a measurement further by automating the clicking and screen-scraping. The existent worth is successful connecting models to each different truthful they tin person conversations—so that a exemplary that predicts the terms of maize tin observe upwind forecasts for the coming year. We tin physique networks of AI models and agents. That’s what MCP supports. We couldn’t ideate this exertion conscionable a fewer years ago. Now we can’t conscionable ideate it, we tin commencement gathering it. As Blaise Agüera y Arcas argues, quality is corporate and social. MCP gives america the tools to physique artificial societal intelligence.

The manufacture has been talking astir agents for immoderate clip now—dozens of years, really. The astir caller burst of agentic treatment started conscionable implicit a twelvemonth ago. For the past twelvemonth we’ve had models that were bully enough, but we were missing an important portion of the puzzle: the quality to nonstop discourse from 1 exemplary to another. MCP provides immoderate of the missing pieces. Google’s caller A2A protocol provides much of them. That’s what discourse serialization is each about, and that’s what it enables: networks of collaborating AIs, each acting arsenic a specialist. Now, the lone question is: What volition we build?