Around Christmas 2025 many AI engineers noticed agents ‘‘started to work.’’ The improvement isn’t attributable to a single, isolated change: part of it was simply having time to try new agents with new models, but crucially the models and the surrounding systems—the agent harnesses—had both advanced and reached a favorable intersection. The main argument here is that co‑improvement of models and harnesses created the observable leap: models began absorbing harness capabilities into their weights, engineers deleted what was absorbed, and the remaining harness functions shifted toward mediating human attention.
Defining the agent harness
An agent harness is everything outside model weights that lets an agent operate: environment, tools, context, guardrails and permissions. Without it a model is a ‘brain in a vat’—it can take an epistemic decision but needs the harness to actuate that decision in digital space. The harness gives the model a body: perception (context), action (tools), persistence (memory and compaction) and boundary enforcement (permissions and guardrails).
Harness 1.0 — the early, bolt‑on era
Model and harness evolution can be seen as two curves: what the harness asks of the model, and what the model can actually deliver. The gap between them equals agent effectiveness.
-
ReAct (October 2022): ReAct introduced an ‘‘agent loop’’ as a prompting technique (reason -> act -> observe -> repeat). At that stage the harness was primarily a prompting construct; both curves were near zero.
-
Toolformer (Meta, February 2023): hinted that tool use could be learned rather than only prompted.
-
AutoGPT / BabyAGI (Spring 2023): harnesses sprinted ahead, handing models full autonomy while the models remained brittle next‑token predictors. Loops amplify existing capabilities — and below a threshold they amplify errors. For example, 95% per‑step reliability over a 20‑step task can yield only ~36% average success. This period was the widest gap.
-
Cursor / Copilot (2023–2024): early AI‑powered IDEs responded by reverting to human‑in‑the‑loop designs: don’t give the loop to the model directly, give it to the human and let the model speed the human up. Tests (e.g., by the Answer.AI team) showed that handing autonomy back to the model at that time produced low (~15%) success rates.
Claude Code and the crossing of curves (February 2025)
By late 2024, model capabilities (with the first reasoning models like o1) began to overtake the harness. That created an opportunity someone had to seize. Anthropic’s Claude Code (February 2025) did that: it moved from IDE integrations to terminal‑level access, giving the model bash and file read/write permissions and replacing per‑change human approval with permission rules. This time the model understood the assignment. Built by Boris Cherny and team with the next model’s capabilities in mind, Claude Code grew to roughly $1 billion ARR within six months because it was launched at the crossover point where model reliability finally supported autonomy.
Harness 2.0 — the co‑training era (present)
Today the harness has a measurable effect. Harness‑Bench ran the same model on 106 tasks with different harnesses and found scores ranging from 52.4 to 76.2 — a 23.8‑point spread with zero change to the model. OpenAI reported similar effects on ARC‑AGI‑3: retaining reasoning and compaction uplifted GPT‑5.6 Sol’s score from 13.3% to 38.3%.
Mechanically, reinforcement learning has moved inside the harness. OpenAI’s codex‑1 announcement (May 2025) stated that codex‑1 was trained using reinforcement learning on real‑world coding tasks in a variety of environments. Tool call behavior is no longer only prompted from outside; tool usage is trained within the model’s environment. As models are trained that way, they start absorbing harness capabilities into weights — e.g., learning to auto‑compact with awareness of their own context window. The GPT‑5.1‑Codex‑Max launch was described as the first model natively trained to operate across multiple context windows through compaction.
Once a capability moves into the model, the harness can remove scaffolding. Thariq Shihipar at Anthropic reported deleting about 80% of Claude Code’s system prompt. The metric of harness evolution becomes how much of the harness you can delete while keeping the same capability level. The iterative loop is: train -> absorb -> shed -> repeat.
Harness 3.0 — the Attention era (future)
If the model eventually absorbs every computable harness capability, what remains? The author argues that human‑centric agent functions will persist: permissions, identity, trust and legibility. Absorption inverts rather than ends the harness: it becomes the model’s interface to human attention.
As Ryan Lopopolo put it on the Latent Space episode “Extreme Harness Engineering for Token Billionaires”: “The only fundamentally scarce thing is the synchronous human attention of my team.” Tokens have become abundant and reliable, but human attention remains the bottleneck.
The prediction is that within a year, companies building agentic AI will ship a human attention policy surface — the ‘‘attention‑interface’’ analogous to AGENTS.md (which tells an agent how to work with your codebase). The attention‑interface will specify when the agent may interrupt, when it should continue independently, which decisions it can make alone, and which require human approval. Like other components, it will be learnable: every human correction will provide training data.
Conclusion
The model started as a brain in a vat; the harness gave it a body; the body is progressively dissolving into the brain. What remains to design are the interfaces that future models cannot fully absorb — interfaces that govern human attention, permissions, identity and trust. Those interfaces will determine how next‑generation agents integrate into human workflows and decision processes.
