Coding agents went from unreliable to indispensable over a single winter, and the trigger was not a new model card. Dan McAteer, writing for Latent Space, argues the jump came from two separate curves finally lining up: how much autonomy the surrounding software handed a model, and how much autonomy that model could actually handle without falling apart.
McAteer’s case rests on a history of mismatches. ReAct, the October 2022 prompting technique that defined the reason, act, observe loop, existed purely as a method for coaxing a model to behave like an agent; it changed no weights. AutoGPT and BabyAGI, released the following spring, handed models full autonomy that the underlying next-token predictors could not support. McAteer cites the math: a model with 95 percent reliability per step, given a 20-step task, succeeds only about 36 percent of the time. The scaffolding outran the model, and the result was compounding failure.
The correction, per McAteer, was Cursor and Copilot pulling the loop back toward a human operator instead of the model, a retreat he calls the right call rather than a failure of nerve. He notes that engineers at Answer.AI benchmarked early Devin and found its success rate near 15 percent, evidence the industry was not ready to hand the loop back yet. That changed once OpenAI shipped o1 in late 2024, the first reasoning model, which McAteer says inverted the gap for the first time: model capability now exceeded what harnesses were asking of it.
Anthropic’s Claude Code, released in February 2025 under Boris Cherny’s team, is where McAteer locates the actual crossover. It moved coding assistance out of the IDE and into a terminal with direct file and shell access, replacing manual approval on every step with permission rules instead. McAteer credits its roughly $1 billion in annualized revenue within six months to timing, not novelty: it was built for a model that had already caught up.
From there, McAteer says the two curves stopped meeting and started merging. He points to Harness-Bench, which ran one unchanged model across 106 tasks in different harnesses and recorded scores from 52.4 to 76.2, a 23.8 point spread attributable entirely to the scaffolding. He cites a comparable OpenAI result on ARC-AGI-3, where adjustments to reasoning retention and compaction alone tripled GPT-5.6 Sol’s score from 13.3 to 38.3 percent. As reinforcement learning moved inside the harness (McAteer references OpenAI’s codex-1, trained on real coding tasks) models began absorbing skills like context management directly into their weights, and Anthropic’s Thariq Shihipar has said the team deleted 80 percent of Claude Code’s system prompt as a result.
McAteer’s forecast is where the piece earns its space in this newsletter. As models keep absorbing what the harness used to do, what remains is not code orchestration but the interface between an agent and its human operator: permissions, interruption timing, and what decisions require approval. He predicts that within a year, agentic companies will each publish what he calls an “attention policy surface” alongside their AGENTS.md files, a document governing when the agent may interrupt a person versus keep working unsupervised.
That reframing matters more than the history lesson. If the harness is no longer competing to keep pace with the model but with a person’s limited attention, the product being sold is not autonomy, it is judgment about when to interrupt. Teams building agent products should treat interruption logic, approval routing, and escalation rules as core differentiators, not afterthoughts bolted onto a capable model.
Dan McAteer, “The Evolution of the Agent Harness,” Latent Space, published August 22, 2026.
