- Author
- Bassim Eledath
- Published
- 10 March 2026
- Kind
- Personal blog post, a practitioner framework from experience and conversations
- Original
- www.bassimeledath.com/blog/levels-of-agentic-engineering
- Subject
- Harnesses and agent loops
- Themes
- Handing over control
What it is
Bassim Eledath wrote this post on AI-assisted coding. It names eight levels and describes what each involves. It argues that capability has outrun practice and that teams should climb the levels together.
What it says
- The post sets out eight imperfectly sequential levels, from tab complete to agent teams.
- Eledath says context engineering shifted from filtering bad context to supplying the right context.
- Compounding engineering is a plan, delegate, assess, codify loop.
- Backpressure such as tests and linters lets agents self-correct. Without it, Eledath says, output is slop.
- Eledath predicts plan mode is dying, but only after levels 3 to 6 are done.
- A separate model should review, since a model grading its own work is biased.
- Eledath calls level 8 an unmastered frontier and favors level 7 for most work.
What evidence it gives
The post relies on the author's experience and on secondhand accounts. Examples include a PR review skill, a CLI called converse, and Dispatch. It cites reports from OpenAI, Anthropic and Cursor but gives no measurements.
Interests and limits
The levels are the author's own framework and are called imperfectly sequential. The author built the Dispatch skill and promotes it. Level 8 claims rest on public reports by others. No data tests the claimed gains.