AI autonomy level framework

Cyrus Radfar on originally posted on LinkedIn
4 min read
ai , autonomy , frameworks , development

AI autonomy level framework

Background

It’s time we have a leveling framework for looking at AI Autonomy outside of self-driving, right?

Frameworks provide:

  • A shared way of thinking and tracking about how you (and your org) should adapt based on the level
  • A shared language to define where we are on a progression so we can evaluate a space

One other idea that we’ll unpack is related to another term that we’ll appropriate from 3D characters and gaming — the “Uncanny Valley”.

We’ll discuss a similar phenomenon that could describe where we are with AI our comfort and trust for AI coding in product development.

AI is different because it’s not just a technology, it’s a technology that’s coupled with a methodology. Therefore, managing the change is fundamentally different. Teams need to learn to dance with their new partner because it’s moving very quickly and no team or organization can get comfortable.

Defining the Levels

Each level of the technology is coupled with a larger change than historically for teams. Adoption requires the technology upgrade (under the hood) and the methodology & process update to adopt the new tools into a process.

In the end, the more autonomy AI requires, the more the organization needs to change.

Here’s my perspective “levels” of autonomy applied to product development.

0️⃣ Level 0: Direct AI Usage - Copy and Paste (2022 - 2024)

Transition enabled by models having large language ChatBot Model access.

1️⃣ Level 1: AI in IDE’s accomplishing tasks - AI makes edits with “tools” (2024-2025)

Transition enabled by models having “Tool Use”

This phase is where many of us were and some still are, working directly with Chat bots vs within tools where we collaboratively build and manage context.

2️⃣ Level 2: Agentic AI in IDEs accomplishing workflows - AI builds for short times with “Tools” (2025-2026)

Transition enabled by models acting Agentically with more tools: Memory, Task Management, Web use, etc.

This phase is where many are today as they migrate to Vibe tools like Cursor and Windsurf that are living in the IDE.

3️⃣ Level 3: Agentic AI in Meta Tools - AI builds continuously behind deeply technical tools. (2026-?)

Transition enabled by human trust in Agents and empowering them to plan more along with improvements in planning, reasoning, etc.

These tools are being built now but haven’t become mainstream. These tools are fully Agentic, run continuously and self-improve. They still require human intervention but are built on the fundamental assumption that the people don’t necessarily need to manage the code. In general, these tools will only work on new projects and products vs on existing repositories. So only new companies or companies that throw out legacy products and rebuild with Agents will benefit from this level of autonomy.

4️⃣ Level 4: Agentic AI as a Department - AI builds and acts as a managed org with executive guidance from people (2026-?)

Transition enabled by orchestration technologies and improved models that are orchestration aware.

The step from level three to four is more subtle than 2 to 3 because the Meta Tools can naturally evolve to get higher level. Level 4 is highlighted by how you manage not what you manage. Managing with high level plans, strategies, and communicating with “AI Leadership” that delegate to swarms of agents is what we’ll see as these products gain adoption. When I speak to my own work on “Vibe Leadership” — I’m betting on this reality.

5️⃣ Level 5: Full Autonomy: Humans out of the Loop, AI decides what to build and releases on its own. (2027 - ♾️)

Transition enabled thanks to improved intelligence (near AGI) or the advent of Swarm Intelligence

Where Are We Today

Today, AI coding feels like chatting with a prodigy who keeps nodding off mid-sentence — brilliant flashes followed by bewildering blanks.

We’re experiencing in that “Uncanny Valley” where generated code looks right, compiles even, but collapses under real-world edge cases.

Here’s the good news: this is a temporary glitch in the growth curve. Give it 6–12 months and we’ll move from Level 2 to Level 3.

Why I’m optimistic:

  • Context windows are exploding. Soon, models will read your entire repo—not just the last few thousand lines.
  • Fine-tuning is getting cheaper. Domain-specific copilots will know your stack as well as your senior devs.
  • Tooling is maturing fast. Linters, test generators, and observability layers are being rebuilt around AI-first workflows.
  • Orchestration software for Agents is in its nascency but will prototype and model the future of a system of Agents as a whole department or organization.

What This Means For Builders Today

  • It’s OK to be skeptical but not cynical. We’re in the “Uncanny Valley” but this too shall pass. AI work may feel junior or mid-level now but plan for the fact that it’ll be senior in 2026 and beyond.
  • Invest in context/prompt hygiene and repo hygiene. Clear structure beats clever hacks when context selection still stumbles.
  • Capture learnings publicly. The playbooks you write now will become your moat when the valley flattens out.
  • I’ve stopped asking, “Can the model do this?” and started asking, “What guardrails and tests let the model fail safely while we learn?”

Keyboard Shortcuts