Write your agent harness like a lease: most of it has an absorption date

Absorption is not uniform, and the uneven part is where the planning happens. The harness, as latent.space defines it, is everything outside the weights that lets a model perceive through context, act through tools, persist through memory and compaction, and enforce its boundaries through permissions and guardrails ^[5]. The first three describe work a model could plausibly learn to do inside itself, and the route was sketched early: Toolformer, from Meta in February 2023, suggested tool use could be trained in rather than prompted ^[6]. Permissions are a different animal. They encode what the person holding the credentials is willing to lose, and no amount of post-training settles that on the model's behalf.

The compounding argument in that history has a second half the essay leaves on the table. latent.space's position is that a loop amplifies whatever capability the model already has rather than adding any ^[8], which means a harness author has two levers and owns only one of them. Raising per-step reliability is the lab's job. Cutting the number of steps is yours. Hold per-step reliability where it was and shrink a twenty-step task to five, and average success moves from roughly 36 percent to about 77 percent ^[1]. Push the other way and ask for 90 percent end to end across twenty steps, and the step needs to succeed 99.5 percent of the time ^[2]. That is a weights number. You do not reach it by wrapping retries around a step that fails one time in twenty.

Which is why the decomposition work, the part that turns twenty steps into five, is worth more than the loop it runs inside. Same for the surfaces that put a person in front of the result. The 2023 to 2024 phase that latent.space labels the retreat to human in the loop, the Cursor and Copilot period ^[11], produced the interfaces that are still in daily use, while the full-autonomy scaffolds of spring 2023 are gone ^[10]. Human attention has survived one cycle of this already. Attention is also the input that does not get cheaper when the model gets better, because a model with more capability generates more output that somebody has to accept or reject.

The schedule is not fast. ReAct put the reason-act-observe loop on paper in October 2022, as a prompting method, before anyone called it a harness ^[9], and it took about 38 months from there to the point where engineers noticed agents working ^[3]. So scaffolding written this quarter will probably earn its keep for a while. The rule is to stop paying for it twice: write it plainly, do not build abstraction layers to make it survive three model generations, and expect to delete rather than migrate.

The whole plan rests on a trajectory latent.space asserts rather than demonstrates: that models keep pulling harness functions into their weights and engineers keep deleting what got pulled ^[4]. If labs prioritise differently, the bet is wrong. Note the asymmetry in being wrong, though. Betting on absorption costs you some code you wrote quickly and threw away. Betting against it costs you a framework you maintained for years around a capability that shipped in a model update.