How fast AI agent techniques age: a timeline
By Philipp Kant 6 min read
AI agent techniques are the methods, file conventions, and workarounds people use to make language models do real work: how they phrase tasks, how they loop an agent, where they keep instructions, how they connect tools. Between 2024 and late 2026 most of them were replaced within six to twelve months. The principles underneath held.
We learned this on our own blog. Our post on the Ralph loop, a popular
way to keep a coding agent working, was written and ready in October 2026.
A check of the last 30 days showed teams deleting their Ralph scripts that
same month, because Claude Code and Codex had built the idea in as
/goal. The post became the /goal post
before it went live.
Two years of agent techniques on one timeline
| When | What appeared | Where it stands in October 2026 |
|---|---|---|
| Sep 2024 | Reasoning models (OpenAI o1) | Standard. “Think step by step” and similar prompt tricks are built into the model |
| Nov 2024 | Model Context Protocol (Anthropic) | The common way to connect tools and data to an agent |
| 2024 | .cursorrules and other per-tool rule files |
Merged into AGENTS.md |
| Feb 2025 | “Vibe coding” (Karpathy), Claude Code preview | In January 2026 Karpathy reported writing 80% of his code through agents |
| Apr to May 2025 | Codex CLI and cloud agent, Claude Code generally available | The two main coding harnesses |
| Jun 2025 | “Context engineering” (Tobi Lütke, Karpathy) replaces “prompt engineering” | Holds: the skill is choosing what the model sees, not the wording |
| Jul 2025 | Ralph loop (Geoffrey Huntley): an agent in a Bash loop | Retired in favor of /goal |
| Aug 2025 | AGENTS.md, by OpenAI, Google, Cursor, Sourcegraph, Factory | The standard instruction file for coding agents |
| Oct 2025 | Agent Skills (Anthropic), open standard since Dec 2025 | Adopted across tools; SKILL.md is a common format |
| Jan 2026 | Personal agents go viral (Clawdbot, renamed OpenClaw within days) | A category of its own |
| Feb 2026 | “Harness” becomes the word for the software around a model | Holds; VS Code’s chat now lets users pick the harness |
| Mar 2026 | Autoresearch (Karpathy): an agent loops on a measured score | The pattern spread to code, prompts, and benchmarks |
| Apr 2026 | LLM wiki (Karpathy); Google’s guide to agent-friendly websites | Both current |
| Apr to May 2026 | /goal in Codex (30 April) and Claude Code (11 to 12 May) |
Replacing hand-made loops |
| Jun 2026 | Lighthouse “Agentic browsing”; WebMCP in Chrome origin trial | Early, still changing |
Most tricks end up inside the product
The same sequence repeats. Someone finds a workaround for a model or tool weakness, the community copies it, a vendor ships a plugin or convention, and within a year it is a built-in feature or a standard.
- Loops. Ralph’s Bash loop (July 2025) became a Claude Code plugin with
a Stop hook, then
/goalin Codex and Claude Code (spring 2026). Ten months from one-liner to product feature. - Instructions.
.cursorrules,CLAUDE.md, and other per-tool files (2024) converged into AGENTS.md (August 2025), a format several vendors agreed on. - Prompt craft. Chain-of-thought phrasing and long prompt templates lost most of their value once reasoning models did the thinking by default.
- Tool glue. Hand-written adapters for each model and each API gave way to MCP servers that any harness can use.
- Know-how. Prompt libraries turned into Agent Skills: a folder with instructions and scripts that an agent loads when the task fits.
The lesson is not that the tricks were wrong. Each one worked, and each one showed the vendors what to build next. The lesson is that a trick is temporary by design.
What aged, what moved in, what lasts
| Aged within months | Moved into the products | Lasts |
|---|---|---|
| Magic phrases in prompts | Loops until done (/goal) |
A written definition of done |
| Bash loops around an agent | Planning and plan modes | Checks that prove the result: tests, validators, evals |
| Per-tool rule files | Instruction files (AGENTS.md) | Small tasks, one at a time |
| Hand-written tool adapters | Tool connections (MCP) | Limited permissions and a human gate by rule |
| Prompt libraries | Skills | State in files, not in the chat |
| Picking “the best model” once | Model switching inside the harness | Clean, owned data |
Everything in the right column was true before agents and is still true now: it is how careful engineering and well-run processes work. The left column is what a model generation made necessary and the next one made unnecessary.
What it means for a company’s AI project
A company that builds on the right column keeps its investment when the tools change. One that builds on the left column rebuilds every few months.
- Spend on what lasts. Data quality, process definitions, the checks that say whether a result is right, the rules for when a person decides. None of it depends on this quarter’s model.
- Keep the model and the harness swappable. The best model changes every few months. An integration that hard-codes one model or one vendor’s workaround is a rewrite waiting to happen.
- Do not pay for glue the vendor ships next. Before a custom build, check what the platform roadmap already covers. A custom loop or memory layer can be outdated before it goes live.
- Review the setup every quarter. What is built in now that was custom three months ago? What can be deleted?
- Check before you build on a technique. A viral post from last spring
is a lead, not proof. We now check the last 30 days before we write
about any AI technique, which is how the Ralph post became the
/goalpost.
How the parts around a model fit together is explained in what an agent harness is; the current way to keep an agent working until a task is done is in /goal in Claude Code and Codex.
FAQ
Is prompt engineering dead? The craft of magic phrasing mostly is. What replaced it is context engineering: deciding what the model gets to see (instructions, data, examples, tools) and checking what comes out. That skill matters more than ever.
How fast do AI agent techniques change? In this period, a popular technique typically went from first post to built-in feature or standard within six to twelve months: the Ralph loop in ten months, per-tool rule files into AGENTS.md within about a year.
Will /goal be outdated too? Probably in its current form. The idea under it, a written end state and an independent check, will not. That is the part worth learning.
How often should a company revisit its AI setup? Quarterly for the tools and models, much less often for the data, the process rules, and the checks. If a review keeps finding custom code the platform now does by itself, the build leaned on the wrong layer.
Which AI skills stay valuable? Writing a precise definition of done, building checks that prove a result, designing permissions and review steps, and keeping data clean. They were valuable before agents and will be after the next change.
Building AI into your systems on the layer that lasts, with the model swappable on top, is our AI integration for business.
Ask an AI about this post
