Referencing this MCP is Dead; Long Live MCP! (via Hacker News), I wanted to highlight what I think is the key principle behind this that will be true for the future of working with AI Agents:
You Need Deterministic Gates for Probabilistic AI Agents
The referenced article addresses the current zeitgeist that favours agents using CLI directly rather than MCP (or other less structured means of calling APIs).
Agent Skills, AGENTS.md etc are all tools for improving the behaviour of LLMs, whether that be via increased skill or greater predictability. But at the point that the agent is still in control, outcomes are still probablistic. We forego guarantees around behaviour so we can benefit from the creativity of AI.
If we want guarantees, (and we do when it comes to security and correct processing of data), we need deterministic gates. The most robust, best specified means of achieving this is MCP, and currently, the best way to manage this in teams is via remote MCP over HTTP.
This is why I'm using remote MCP over HTTP for anything that matters. Yes MCP can add some overhead compared to direct CLI usage, but you can also build in optimisations that deterministically reduce the response size from remote APIs and therefore reducing context bloat. It's also a lot easier to share remote MCP servers, just like the internet helped us share software.
For an example of a light weight MCP over HTTP implementation, see my example of a custom MCP server for Google Workspace integration.
I'm happy for the the agent to use its intelligence and creativity to help me solve problems, but for a range of operations, I want a gate past which actions run with the certainty of normal software functions. This might be so that I can enforce scoped operations rather than give an agent access to everything an API offers, or I just because I want to be sure about what data is read or changed and how. There are many safety critical applications ins software that require the same.
Other Tools Moving in a Deterministic Direction
Other tooling around AI we're seeing evolve around Agentic AI is adding deterministic gates that the probabilistic AI agents work with.
NanoClaw is an example of an application that sells itself on using deterministic filtering of your WhatsApp messages before the agent gets to see them, and proxies API keys so the agent never gets them - this is a similar type of deterministic gate that allows for more confidence when working with AI.
Summary
Certainty around fundamental principles in software such as security and correctness of process requires determinism. AI Agents while improving will never be 100% deterministic (that's what code's for!).