The Agent Harness: From Autonomous Loops to Human Attention

Autonomy needs reliability

In spring 2023, AutoGPT and BabyAGI handed models autonomous loops: take an assignment, act, observe and keep going. Their ambition exposed a reliability problem. A model that often chooses a useful next action can still struggle to complete a long task. In a hypothetical sequence where each of 20 required steps has a 95% conditional probability of success, the probability that every step succeeds is 0.95²⁰, or about 36%. Iteration can compound errors as readily as progress.

An agent harness is the software outside a model’s weights that lets it operate: its environment, tools, context and guardrails. - Context supplies information, tools execute actions, and memory preserves information across a run. - Compaction reduces accumulated context so useful information can carry forward. - Permissions enforce boundaries on what the agent may do.

The central design problem is matching what the harness demands to what the model can reliably deliver. Too much autonomy exposes failures; too little leaves useful capability idle. This offers an explanation for the perceived jump in agent usefulness around Christmas 2025: models and harnesses improved together. Lukasz Kaiser described the jump as difficult to isolate because harness changes, post-training changes and new pretrained models arrived around the same period.

Giving the model control

The loop preceded those early autonomous products. ReAct, introduced in October 2022, used prompting to make a model reason, act, observe the result and repeat. The procedure came from outside the model. Toolformer, introduced in February 2023, pointed toward another approach: training models to use tools. That distinction—externally instructed behavior versus learned behavior—runs through the evolution of the harness.

After AutoGPT and BabyAGI pushed beyond what contemporary models could reliably sustain, Cursor and Copilot offered a more constrained arrangement in 2023–2024. Humans orchestrated the work while models accelerated individual actions. Early Devin attempted to return more autonomy to the model. An Answer.AI test cited in the article reported roughly 15% success, though the article provides neither a task count nor evaluation conditions for interpreting that result.

The proposed turning point came near the introduction of o1 in late 2024, when model capability began to exceed what constrained interfaces asked of it. Claude Code exploited that opening in February 2025 by giving the model a terminal, bash and file access. Permission rules replaced approval for every change, allowing the model to run more of the loop itself.

Boris Cherny’s team reportedly designed Claude Code around the next model’s capabilities. The article reports roughly $1 billion in annual recurring revenue within six months and interprets its growth as a result of well-timed autonomy. That causal connection remains an interpretation, but it captures the engineering bet: build for capabilities becoming reliable enough to use.

Training inside the harness

The surrounding system can materially change performance without changing the model. - Harness-Bench reportedly evaluated the same model on the same 106 tasks across different harnesses, producing scores from 52.4 to 76.2—a 23.8-point spread. - In another reported result, adding retained reasoning and compaction raised GPT-5.6 Sol’s ARC-AGI-3 benchmark score from 13.3% to 38.3%. Keeping earlier reasoning available and carrying context forward changed what the agent could accomplish in that evaluation.

Training within the operating environment brings model and harness development closer together. OpenAI’s May 2025 codex-1 announcement described reinforcement learning—learning from feedback on task performance—on real-world coding tasks across multiple environments. Tool use could then become behavior learned through those tasks, reducing reliance on instructions added after training.

This creates a path for models to learn coordination previously supplied by external scaffolding. The GPT-5.1-Codex-Max launch description, for example, said the model was natively trained to operate across multiple context windows through compaction. Maintaining continuity when a task exceeds the context available at once becomes partly a learned capability.

The resulting engineering cycle is to train a capability, let the model learn it, remove redundant scaffolding and repeat. Thariq Shihipar reportedly said Anthropic had recently deleted 80% of Claude Code’s system prompt. The proposed measure of progress is how much harness code or instruction can be removed while preserving capability. Deletion matters because the behavior survives without that support.

Permissions remain outside the model

Multi-agent orchestration—the coordination of several agents—tool selection and memory are possible next targets for learned behavior. As more of this coordination moves into model weights, the harness may need less scaffolding to help the model operate a computer.

Permissions impose a different requirement. A model can learn to respect a boundary, but permission enforcement must retain authority outside the model. Treating permission as something the model merely remembers dissolves its role as an external constraint. Permissions, identity, trust and legibility—making the agent’s behavior understandable to its operator—remain human-facing responsibilities of the harness.

The harness therefore changes emphasis rather than disappearing. As models handle more computer-facing work, the surrounding software increasingly manages their relationship with people. The mismatch shifts from what the harness asks of the model to what the agent asks humans to decide, and what those humans have the capacity to answer.

An interface for human attention

Ryan Lopopolo identified his team’s synchronous human attention as its fundamentally scarce resource. Even autonomous work can depend on someone stopping to inspect a situation and make a decision. Long-running agent progress files and approval queues are early examples of interfaces that help manage that demand.

The closing proposal is a human attention policy, analogous to AGENTS.md, a file that gives agents instructions for working with a codebase. An attention policy would tell the agent how to work with its operator: when it may interrupt, when it should continue independently, which decisions it can make and which require approval. Writing in August 2026, the article predicted that every company building agentic AI would ship such an interface within a year.

That interface could learn from corrections. Feedback about unnecessary interruptions or decisions that needed approval would become data for adapting how the agent works with its operator. As models absorb more operational coordination, the enduring harness problem becomes how they request attention and authority from the people responsible for them.