The new model was already in the room.
The shared experiment had exposed the immediate problem: giving a stronger model a seat at the table did not make the table coherent. Dell Manager had no settled shape. The existing agents carried different histories and assumptions. Everyone could produce; nobody agreed who should lead.
The experiment had changed the question. The operator was no longer asking only whether the model was stronger. He was asking what happened when new behaviour entered an operating environment built around older models.
The existing system ran on a small machine at home. A more powerful retired laptop was already being prepared as its possible replacement. That hardware story belongs to the next part. The problem here was the adaptation tax.
A model release looks wonderfully clean from the outside. New capabilities. New context claims. New benchmark charts marching across the internet like tiny victorious soldiers.
Inside a real agent system, however, a new model is not arriving alone.
It walks into a house full of old instructions.
Ana had prompts written around previous behaviour, memories accumulated under previous assumptions, skills carrying old procedures, guardrails designed to catch yesterday’s mistakes, compression settings, tool schemas, routing habits, and enough historical context to make a fresh model feel like the newest employee in a company where nobody has cleaned the shared drive.
The first responses could look excellent.
That made the later drift harder to trust.
The operator would give the model a goal. It would appear to understand. Work would advance. Then something changed: an element he had not asked to change, a confident reinterpretation, a simplification that weakened the original requirement, or a later turn that no longer seemed anchored to the beginning.
So he corrected it.
The conversation grew.
The correction joined the earlier prompt, the tool results, the summaries, the files, the new rules, and the residue of whatever the harness had already decided mattered. Eventually the context needed compression. The system kept the shape of the work while parts of the original intention lost force.
This is not proof that GPT-5.6 is bad.
It is not proof that compaction caused every failure either.
It is evidence of something less marketable: a new model and an old operating environment do not automatically fit.
The human had learned how to work with earlier models. Not perfectly, but enough to build reflexes. He knew when to explain more, when to add a guardrail, when to demand verification, and where certain failures tended to hide.
Then the behaviour changed.
Now the same detailed prompt could become noise. The same freedom could become reinterpretation. The same protection could compete with the task. A repair that had once prevented drift could help produce it.
This is the adaptation tax.
It is paid in prompt rewrites, test conversations, changed boundaries, broken assumptions, new verification, and hours spent discovering which old lessons remain true. The bill arrives before the new model has saved anything.
And the tax repeats.
The operator described the cycle plainly: learn the model, tune the workflow, fight through the hallucinations, stabilize it — then expect another model in a month or two and begin again.
The cruel part is that the model may genuinely be better.
Better reasoning does not remove accumulated contradiction. More context does not guarantee that the live objective dominates stale material. Stronger agency does not make the surrounding harness coherent.
The model is the engine.
The prompts, memory, tools, policies, skills, state, and stop conditions are the vehicle.
We had changed the engine and expected the old vehicle to become new around it.
By the end of the experiment, the operator was not asking which model had won.
He was asking whether another patch would leave him rebuilding the same agents again when the next model arrived.
The answer was not another guardrail.
The possible answer was already sitting beside him: the retired laptop he had spent the week preparing. The next question was no longer how to tune the new model inside the old environment.
It was what, exactly, deserved to move into the new one.