23 · TL;DR in 5 points

Why it matters

The harness is starting to be as important a piece as the model, and now there is one you can read end to end.

So far, what it is and how it is built. What is left is why it should matter to you.

For whoever builds product

  • Old option From scratch Months of work on things that are not your product
  • Old option Somebody else's harness And finding out halfway that the piece you need to change is exactly the one you cannot touch
  • New option A replaceable base Any piece, including the one that decides what the agent does, swaps out without forking the project

The cost is not zero. It is reading a whole composition framework. But it is a cost you pay once, instead of paying it every time you hit a limit.

For open models

An open model without an open agent layer only gets you halfway. You can download the weights, serve them wherever you want, and then you need a product on top that is almost always closed.

With both layers open, the full chain stops having a stretch you do not control. For anyone with sovereignty or audit requirements that is not an incremental improvement: it is the difference between being able to and not.

For whoever runs infrastructure

The session log is the least flashy part and probably the most useful one here. Everything the model saw is written down, so you can reconstruct why an agent did what it did.

When someone has to answer to a customer, an auditor or a regulator for what an agent did, that reconstruction stops being a convenience and becomes a requirement.

For the industry

A lab that sells inference opening its agent layer under an MIT licence and making it compatible with competitors’ models is a counterintuitive decision.

What is checkable is the effect. From now on, a closed agent product has to justify why it is closed.

TL;DR

  • One An agent is a model plus a harness The model sets the ceiling, the harness decides how much of that ceiling reaches your machine.
  • Two Same model, different results Two products with the same model behind them behave differently, and the difference is in this layer.
  • Three DeepSeek published theirs Under an MIT licence, compatible with third party models, and with an 88 page paper formally proving why its architecture survives being changed at runtime.
  • Four There is no privileged core It is the first of its kind where the agent loop is replaced by configuration. Every piece is a plugin alongside the others.
  • Five It is a preview And it will break compatibility. Worth understanding now and committing to later.