From prompts to ecosystems: The new rules of AI
The transition from prompt engineering to harness engineering signals a fundamental shift: the focus is no longer just on how we talk to AI, but on the operational environment housing it. Organizations must stop obsessing over which base model to activate and instead strategically design the systems, constraints, and feedback loops that surround these intelligent agents. Ultimately, as models rapidly converge in baseline capability, your competitive advantage will belong to the robustness of your AI infrastructure rather than the cleverness of your prompts.
Points clés
- According to Sayash Kapoor’s HAL benchmark, a programming task’s success rate surged from 42% to 78% simply by changing the AI’s environment to Claude Code, using the exact same model and data.
- The evolution of AI interaction shifted from Prompt Engineering in 2023-2024 to Context Engineering in 2025, a concept highly popularized by researcher Andrej Karpathy.
- The tech community recognizes 2026 as the era of Harness Engineering, defined by companies like Anthropic and OpenAI as the complete structural framework surrounding a model.
- Researcher Philipp Schmid compares the harness to a computer’s operating system, which orchestrates the raw processor (the AI model) and its memory (the context).
- An OpenAI software team of up to seven engineers utilized a Codex-powered agent to build a one-million-line application in five months, with zero lines written by a human.
- Reflecting on their success at OpenAI, engineer Ryan Lopopolo noted that managing the agents was not the bottleneck; building the harness was the real challenge.
- Stripe’s internal autonomous agent system, named “Minions”, autonomously merges over 1,300 code modifications weekly without humans writing a single line.
- An industry debate is currently unfolding between the “big model” advocates, such as Anthropic’s Boris Cherny and OpenAI’s Noam Brown, and the “big harness” proponents like Jerry Liu from LlamaIndex.
- Mitchell Hashimoto, creator of Terraform, asserts that engineering a harness means building an infrastructural system to prevent an AI from repeating an error, rather than simply correcting its prompt.
- As titans like Google, Microsoft, and Notion converge toward fundamentally similar AI architectures, organizational differentiation relies entirely on proprietary context and agile feedback loops.
À retenir
If you are still pulling your hair out trying to whisper the ultimate, game-changing prompt into ChatGPT’s ear, congratulations on playing a 2023 game in 2026. It is time to treat AI less like a magical oracle and more like an intern swinging a sledgehammer—meaning you actually need to build a workspace with some solid guardrails. Stop endlessly shopping for the trendiest new model and start engineering the operating ecosystem around it, because handing a supercomputer to someone with no operating system and expecting a masterpiece is definitely a bulletproof business strategy, right?
Sources
