Shipping the five-step agent loop costs between 794 and 1,729 lines

Five steps describes the loop correctly, but the description does not size the work of building one. Assemble context, call the model, execute any tool calls, append the results, decide whether to go round again ^[4]. Claude Code's `while(true)` opens at line 307 and closes at 1728, so the body alone is 1,421 lines ^[5]. That averages 284 lines per step ^[2], with a further 308 lines of the same file sitting outside the loop ^[3].

The extra bulk is plumbing: streaming, where tool calls arrive before the message is finished and you have to decide whether to start executing; parallel execution, which is safe or unsafe depending on what the tools touch; mid-turn interrupts, which have to unwind without corrupting the transcript; compaction when the context fills; telling retryable provider failures from the other kind; and approvals, which can park the loop indefinitely while a human decides ^[8]. None of it is intelligence. A tutorial loop is 50 lines because nobody in a tutorial hits escape halfway through a turn.

The counts themselves travel badly. They come from a single clone taken on 2026-09-01 for a teardown published on dev.to ^[3]: 1,729 lines of TypeScript, 983 of Rust, 794 of TypeScript ^[2], with Codex and Pi spreading comparable work across neighbouring modules, which is why the author compares file to file rather than loop body to loop body ^[6]. The 189 lines between Codex and Pi ^[5] tell you about module boundaries and brace style, not about difficulty. The factor of 2.18 across the full range ^[1] survives translation because it is coarse. Measured against the "~50-200 lines" row the same author printed earlier and has now retracted, Pi alone is roughly 16 times the low end ^[4][7].

The divergence worth reading is the boundary: what sits outside the loop, and how the loop talks to it ^[9]. Claude Code keeps the loop, the tools, the permission checks, compaction and the UI in one process, coordinating by function call and shared state, which is what lets its five-level compression pipeline decide using information the loop has not committed to anything yet ^[10]. The bill for that is reach. Everything must be in-process, and another program cannot drive the agent except through the surface Anthropic chose to expose ^[11]. Codex went the other way, pulling the agent core into `app-server` behind a documented, bidirectional JSON-RPC 2.0 interface whose README starts from a Thread primitive, one conversation between a user and the agent ^[12].

In my context that decision comes before the line count. If you own the process and the UI, in-process wins on compaction fidelity, because the compressor can use state that never had to be serialised onto a wire ^[10]. If you want a job runner or an IDE to drive the agent, the RPC surface is the part you cannot retrofit cheaply ^[12]. And a harness well under 800 lines has made a choice about which of those six hard parts it does not handle, which is worth establishing before you adopt it.