GPT 6 Astra, A Tale of Two Harnesses

Astra was fighting against the harness. That is really what this post was saying. I don't have measured proof of the claim, but enough from my own experience to decide I won't be wasting time and tokens going down that path again.

Now, I could make this claim with authority. That is the style of the day after all, to make a bold assertion and then back it up with rephrasing of the press release about the product the claim is being made on. I see how it works, and I reject it.

So what I will say is this: I can't tell if OpenAI deliberately nerf models when they are in a third-party harness or not. I suspect it is more a case of fine-tuning the model to be optimised in their proprietary harness, a completely reasonable path for a commercially motivated company to take. There is no doubt the user experience is much better for GPT models in Codex than I have found using them in Hermes.

I spent a weekend battling the pedantry of Astra in Hermes. Asking it to do similar tasks in Codex met with far greater success, no pushback, and the confidence to trust it with significant work. And therein lies the rub, whether it is better to be constrained in an environment dictated by others who "know what's best for you", or to maintain independence and be the master of your own (a word like destiny, but not as grandiose) (1).

All the leading AI companies are pushing their proprietary harnesses. That much is observable. Here is the part I can't prove, and it is exactly the kind of claim I said I would reject: harness lock-in is the way the company strategists see the future.

Here is what I think. AI pundits lauded prompt engineering as the necessary skill for users in 2025. Since early 2026, harness engineering has taken prime position. What is not settled is who writes the harness.

Astra 6 and Fable 5 are so capable, they can write their own harness (with the caveat they do it in a harness they are not fighting against in the first place). It might cost a few hundred bucks in compute, maybe even a few thousand. In other words, the price of a single desktop PC for one employee. In return the organisation gets its own no lock-in harness. It will come with trade-offs, everything does. The upside is that the model capable enough to write the harness is also capable enough to maintain it. And when a competitor leapfrogs it, as Fable leapfrogged GPT 5.6 and Astra leapfrogged Fable, you move to the new leader and it maintains the harness instead. That is what not being locked in buys you.

In fact "Astra, start a project to review my current use of the Hermes and Pi harnesses, compare it with the attached 'Use and Purpose' briefing, and write a harness that will fulfil the criteria in the 'Harness Scope.md' document."

There, done (2).

(1) Clearly a placeholder for me to look up in a thesaurus later, but my editor agent, Clare, thought it was funny as is, so I left it in.

(2) Clare takes this to mean "easier said than done". It is right of course, some iteration will be required.