MCP isn't the USB-C of AI: LLMs improvise, so put a deterministic runtime in charge
ThomasBuildLab · reddit · 2026-09-27
The author pushes back on the popular "MCP is the USB-C of AI" analogy: USB-C is deterministic, but with MCP the orchestrator is an LLM — whether to call a tool, which one, with what arguments, what comes next, and when the job is done are all improvised by the model. That's where the freestyle starts.
He proposes a cleaner layering:
- The LLM only interprets intent from language
- Structured intent feeds a deterministic runtime that executes
- Invariants, state machines, transactions, and validation govern what's allowed
- MCP is demoted to pure transport
His conclusion: the next step in agent infrastructure isn't smarter agents — it's giving them less freedom.
More from coding & agent
- Grok Build v1.0.41 ships subagent config inheritance and long-reasoning fixes — mark_k · 2026-09-27
- Harrison Chase on building frontier harnesses: context layers, evals, online learning — VeryWellVersed · 2026-09-27
- Dev shares public Grok Bot 'chief of staff' template running a background specialist agent swarm — RachelVT42 · 2026-09-27
- Meta pitches hand-first dev workflow: one Unity codebase across Quest and VR glasses — Scobleizer · 2026-09-27
- Antigravity 2.0 ships /plan mode as dev mocks it as behind Anthropic's curve — brandon_galang · 2026-09-27
- Codex power users say Astra burns $200 limits in days while Anthropic's cap is hard to hit — brandon_galang · 2026-09-27