Engineering Agents Instead of Developers: Six Months of a Coding Revolution
The creator of OpenClaw has noticed that agents went mainstream in software development within six months. In May they delivered a wow effect; by December they were just part of the normal workflow. The industry now uses his methods for working with agents.
AI-processed from Habr AI; edited by Hamidun News
Six months ago, Peter Steinberger, the creator of OpenClaw (later acquired by OpenAI), published a post about shipping at inference speed: an AI agent finishes a task in 5 seconds of inference instead of an hour of manual coding. At the time, it read like a revolution and a set of first impressions. Now that revolution is just an ordinary workday across the industry.
From experiments to standard practice
In May 2025, inference speed was an interesting idea. Agents made frequent mistakes, needed custom prompts, and only handled simple tasks. Developers were skeptical: "This isn't real development, it's a toy."
By December, the picture had changed completely. Agents learned to handle real projects with thousands of files, complex integrations, and legacy code. Accuracy rose from 70% to 90%+. Most importantly, iteration time dropped from hours to minutes. The agent spots the error, redoes the work, and pushes the result back.
How it works in practice
Steinberger found a simple truth: delivery speed = inference speed. If the model runs inference in 5 seconds, a feature ships in 5 minutes (minus validation). By comparison, it used to take 1-3 hours.
But speed isn't the main point. The real shift is in the developer's role. You no longer write code. You define the task for the agent, review the result, and either approve it or send it back for rework. It's like hiring a junior developer who never gets tired and works in microseconds.
Methods for working with agents
Over six months, a set of practices emerged that Steinberger described and that are now standard:
- Minimize clarifications — prepare all the context before making a request
- Use the file graph and type hints as training examples for the agent
- Give the agent access to commit history (examples of successful code)
- Don't fix the agent's output by hand — give it the error and it will redo the work itself
- Focus on high-level tasks, not syntax
Engineers now think at the level of features, not lines of code.
The agent takes care of everything else.
What this means for the industry
Development is no longer about writing code. It's now about defining tasks and validating results. Those who have grasped this shift work 3-5 times more productively. Those who still write code themselves are gradually losing their competitive edge.
Steinberger was writing about first impressions, but it turns out they weren't impressions at all — they were the new normal.
- Nov 18, 2026Five AI Models Tried to Fix the Same Bug: Two Succeeded, Three Broke More Than They Fixed
- Nov 18, 2026Developer gives a web AI chat hands: an MCP server with 14 tools and 21,000 lines of code
- Nov 18, 2026OpenAI releases GPT-6.2 Astra: second place in the rankings and a 20% lower price
- Nov 18, 2026Three "token-saving" utilities raised the AI agent bill by 18–46%
Key Questions Answered
What are the key details on From experiments to standard practice?
In May 2025, inference speed was an interesting idea. Agents made frequent mistakes, needed custom prompts, and only handled simple tasks. Developers were skeptical: "This isn't real development, it's a toy." By December, the picture had changed completely. Agents learned to handle real projects with thousands of files, complex integrations, and legacy code. Accuracy rose from 70% to 90%+. Most importantly, iteration time dropped from hours to minutes.
How it works in practice?
Steinberger found a simple truth: delivery speed = inference speed. If the model runs inference in 5 seconds, a feature ships in 5 minutes (minus validation). By comparison, it used to take 1-3 hours. But speed isn't the main point. The real shift is in the developer's role. You no longer write code. You define the task for the agent, review the result, and either approve it or send it back for rework. It's like hiring a junior developer who never gets tired and works in microseconds.
What are the key details on Methods for working with agents?
Over six months, a set of practices emerged that Steinberger described and that are now standard: Minimize clarifications — prepare all the context before making a request Use the file graph and type hints as training examples for the agent Give the agent access to commit history (examples of successful code) Don't fix the agent's output by hand — give it the error and it will redo the work itself Focus on high-level tasks, not syntax Engineers now think at the level of features, not lines of code.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.