Levels of Agentic Engineering
This article outlines eight progressive levels of
The Lowdown
As AI's coding capabilities grow, the true challenge lies in effectively harnessing it. This piece introduces an 8-level framework of "agentic engineering," detailing the progression from basic AI assistance to fully autonomous agent teams. It aims to bridge the gap between AI's potential and its practical application in software development, highlighting how each level unlocks significant productivity gains and why teams should strive to ascend this ladder.
Here are the levels of agentic engineering:
- Levels 1 & 2: Tab Complete and Agent IDE: Starting with basic code autocompletion (like GitHub Copilot) and evolving to AI-focused IDEs (like Cursor) that connect chat to the codebase, enabling multi-file edits. The primary limitation here is context management.
- Level 3: Context Engineering: Focusing on improving the information density of prompts, ensuring the model receives just the right amount of relevant context. This involves careful system prompts, rules files, tool descriptions, and managing conversation history to prevent context overload or underspecification.
- Level 4: Compounding Engineering: Implementing a "plan, delegate, assess, codify" loop where lessons learned from AI interactions are codified into rules or documentation, making future sessions more effective. This is about systematic improvement rather than one-off fixes.
- Level 5: MCP and Skills: Giving LLMs access to external tools and APIs (Multi-Capability Platforms or MCPs) like databases, CI pipelines, or browser testing tools (e.g., Playwright). This allows models to act on the codebase, moving beyond just thinking about it, and encourages sharing and versioning skills among teams.
- Level 6: Harness Engineering & Automated Feedback Loops: Building a comprehensive environment with tooling and feedback loops that enable agents to work reliably without constant human intervention. This includes observability, backpressure mechanisms (tests, linters), and security boundaries, emphasizing designing for throughput and using constraints over explicit instructions.
- Level 7: Background Agents: Shifting from human-in-the-loop planning to agents that can plan and execute asynchronously. This involves orchestrator agents dispatching tasks to worker agents, utilizing different models for specialized jobs, and decoupling implementers from reviewers to avoid bias.
- Level 8: Autonomous Agent Teams: The frontier where agents coordinate directly, claiming tasks, sharing findings, and resolving conflicts without a single orchestrator. While still highly experimental and costly, projects like building compilers or web browsers demonstrate its potential.
- Level ?: The future envisions conversational, voice-to-voice interaction with coding agents, moving beyond text-only interfaces and embracing iterative development in a much faster and more integrated manner.
Ultimately, the article serves as a practical guide for developers to understand where they stand in the AI-assisted coding landscape and how to advance their practices, emphasizing that each step amplifies gains as AI models improve.
The Gossip
Ladder Logic: Critiquing the 'Levels' Construct
Many commenters questioned the article's "levels" metaphor, suggesting it implies a toxic skill hierarchy rather than evolutionary stages of AI-assisted development. A common concern was the idea of "dark factories" at higher levels, with skepticism about whether fully autonomous agents could produce software of sufficient quality or market value without human oversight, or if these advanced techniques are competitive pursuits rather than universally applicable solutions.
Contextual Codification: Beyond the 'What' to the 'Why'
A significant theme revolved around the challenge of codifying not just rules, but the underlying rationale and tradeoffs behind engineering decisions. Commenters argued that current methods, like `.CLAUDE.md` files, are insufficient, and suggested leveraging git history, Architectural Decision Records (ADRs), or structured context blocks to ensure agents understand the "why" and not just the "what," especially for long-term agent autonomy.
Agentic Adventures: Realities of High-Level Automation
Many users shared their experiences with the more advanced "levels" of agentic engineering, particularly 7 and 8. Some reported significant success with custom orchestration layers and multi-agent teams, citing faster development and improved results, particularly with newer models. However, others highlighted the considerable token costs, UI limitations, and operational complexities that make truly autonomous agent teams impractical or uneconomical for most tasks, suggesting the article might even be a few months behind the rapid pace of AI development.
Evolving Engineer: Human Roles in an Agentic Future
Commenters deliberated on the evolving role of the human engineer in an increasingly agentic world. There was skepticism about the utility of voice-to-voice interaction for complex tasks, with many preferring the deliberate process of writing. The discussion also touched on whether engineers continue to learn core programming skills or if their focus shifts primarily to agent orchestration, with some finding themselves learning more than ever through this new paradigm.
Escalating AI: The Humorous Hyperboles of Agentic Ladders
As is common with "level" frameworks on Hacker News, many commenters engaged in humorous speculation, extending the article's 8 levels of agentic engineering to increasingly absurd and dystopian scenarios, envisioning agent managers, CEOs, and even superintelligent entities controlling everything, from the economy to cosmic debates.