Video summary

Forget Everything You Know About Writing Code in 2026

Main summary

Key takeaways

Technology

Main idea: “Loop engineering” (replacing human involvement in coding)

  • The video argues that coding is shifting from prompting AI directly to designing multi-agent “loops” that run with less human supervision.
  • Near-term workflow (next ~6 months): most engineers will spend more of their day on orchestrating/maintaining higher-level automation, rather than writing code via interactive prompts.

Why this is claimed to be “legit”

  • The speaker points out that many AI/job predictions have already come true—especially rapid adoption of AI coding and terminal-based workflows.
  • He references commentary from major tool creators and researchers, claiming they’ve been “right in the past,” though a potential conflict of interest is acknowledged.

Current trend: engineers increasingly rely on AI coding

  • Auto-generated subtitles claim that 90–95% of code at many large companies/startups is now produced via AI prompting.
  • Developers are also said to rely heavily on terminal UI / terminal agents, reducing the need for traditional IDE debugging.

Historical analogy: moving up the abstraction stack

The speaker frames each era as “replacing the previous layer,” such as:

  • punch cards → low-level languages → high-level languages
  • then autocomplete
  • then terminal agents

“Loop engineering” is presented as the next step: replace the engineer’s remaining role by moving from manual prompting to autonomous loop orchestration.

What exactly is an “agent loop” vs “loop engineering”?

Agent loop (current baseline)

A coding agent is described as:

  • calling an LLM
  • iteratively taking actions such as:
    • reading/writing files
    • interacting with GitHub
    • running migrations
    • making git commits/pushes, etc.

This single run is called an agent loop. Limitation (as stated): humans still participate “prompt by prompt,” answering questions or intervening.

Loop engineering (the upgrade)

Add a higher-level (outer) loop that:

  • sets a high-level goal (e.g., “build an e-commerce application”)
  • provides inputs to inner agent loops
  • includes a verifier to check whether the goal was achieved
  • if not, re-runs/continues inner loops until an end condition is met

The outer loop reduces user involvement—the system keeps itself running through orchestration and verification.

“Loops” can be layered many levels

Beyond a single project, examples include:

  • an outer loop that plans/builds/deploys successive products daily (e.g., “new SaaS every day,” “CMS every day”)
  • loops that call multiple specialized components/agents, such as:
    • idea generation
    • execution/building
    • marketing/distribution

Human-in-the-loop (optional safety valve)

The speaker claims humans may still be inserted when:

  • approval is needed for costly actions (e.g., provisioning a paid database)

Otherwise, the thesis is that automation can handle many tasks without constant oversight.

Practical use cases mentioned (examples of loop automation)

The video lists concrete examples of “where this works”:

  1. Documentation loop

    • Claude is described as using scheduled agents to review SDK changes and generate/update/publish docs automatically.
    • A similar approach: when an SDK release is cut, an agent updates docs based on release changes.

  2. Security verification loop

    • A separate agent checks/validates security on every merge/release.
    • Rationale: even if the coding model introduces vulnerabilities, independent verification agents can detect them.

  3. Weekly summaries / newsletters

    • Extract release changes and generate user-facing updates (blog/email/newsletter).
    • May include human approval before sending.

Product/market implications and risks

The speaker predicts:

  • startups will formalize loop systems into products
  • possible consolidation (mergers/acquisitions)
  • unclear definition/standardization right now (a “white space”)

He also warns of a potential wealth gap:

  • those who execute well with AI/loops may see disproportionate growth (revenues, valuations, funding)
  • others may fall behind

Suggested action:

  • evaluate how your job has changed recently, then adapt your day-to-day
  • companies may ask about your AI setup and how you make terminal agents more productive

Main speakers / sources mentioned

  • Andrej Karpathy (cited as having said we’re in a “loopy era,” or similar framing)
  • Creator of Claude Code / Claude (Anthropic) (referenced via claims about agent-driven workflows and documentation updates)
  • Peter (creator of Open Claw; mentioned as acquired by OpenAI)
  • Additional explicit product/source references:
    • Anthropic: two videos on agent workflow:
      • “babysitting your agents or stop babysitting your agents”
      • “build a proactive agent workflow with Claude Code”
    • Cursor and Windsor (named as early autocomplete adopters)

Original video