- Author
- Armin Ronacher
- Published
- June 23, 2026
- Kind
- Personal blog post, opinion based on the author's own use
- Original
- lucumr.pocoo.org/2026/6/23/the-coming-loop/
- Subject
- Harnesses and agent loops
- Themes
- Feedback loops, Handing over control
What it is
Armin Ronacher describes a pattern spreading among people using coding agents. The post sets out where loops work, where they fail, and what follows.
What it says
- The post describes a harness loop outside the agent loop. The harness decides if work is done.
- Ronacher reports little success with loops for code Ronacher cares deeply about.
- Models add local defenses instead of making bad states impossible, and loops amplify that habit.
- Loops work well for porting, performance exploration, security scanning and research.
- The harness only needs a useful signal to continue, not an objective or binary one.
- Opting out may be impossible, because others will run loops against your software.
- The author fears codebases that assume machine participation in their upkeep.
What evidence it gives
The post offers personal experience and argument. It names outside examples, such as a reported Bun port from Zig to Rust. It cites Daniel Stenberg's post on curl. It reports the author's own port of MiniJinja to Go. It offers no benchmarks or measurements.
Interests and limits
The view is a personal taste judgment about code quality. The post says the author wants to understand all shipped code. It discusses Pi, a harness, and the author's changing view of its role. No test compares loop code with hand-steered code.
Connections
These links are this wiki's own reading, not claims the sources make.
- Structural Backpressure Beats Smarter Agents. Ronacher sees models adding local defenses instead of making bad states impossible. Brooks offers types that refuse such states at compile time. But Ronacher's loop needs only a useful signal, while Brooks wants a hard gate. How much signal does a loop need?
- Game design is simple, actually. Koster says rules with no goal make a toy. Ronacher says a harness loop needs only a useful signal to continue. Together they offer a test for any agent loop. Does it have a goal, or is it a toy that writes code?