Summary
Durability is the degree to which an agent harness treats its invocation log as the authoritative state of the world rather than merely a chat transcript. A 2026 harness scorecard ("AI Coding Agent Comparison 2026: 50 Harnesses Scored", kendr.org) evaluated 19 complete harnesses on ten dimensions — loop, edit, ctx, iso, perm, dur, model, ext, surf, cloud — and identified durability as "the one place where the field is broadly weak." Almost every harness persists a transcript so a conversation can be resumed; very few combine single-writer leases, idempotent turn starts, restart reconciliation, and approvals that survive a process kill. The consequence is that the same crash that costs one team a chat history costs another a corrupted working tree.
Key Points
- Durability ≠ transcript persistence. Storing history enables resumption of a conversation; durability makes the log authoritative for execution state so a fresh process can safely continue and reconcile.
- The required ingredients named by the scorecard: single-writer leases, idempotent turn starts, restart reconciliation, and approvals that survive process kill.
- Leaderboard (dur): Kendr Code (10); closest challengers Kiro, Claude Code, Codex at 7.
- Other dimension leaders: loop = Claude Code (10); edit engine = Kendr Code (10); context engineering = Cursor (10); isolation/sandboxing = Codex (10, Seatbelt on macOS / Landlock+seccomp on Linux / network off by default); permissions = Kendr Code (10); model layer = OpenCode and Cline (10); extensibility = Claude Code (10).
- Repository retrieval is a second named structural gap. Cursor and Windsurf maintain semantic indexes over the repo; most terminal-native CLI harnesses rely on agentic search (grep, glob, ranged reads), which is transparent but token-expensive. Aider is the open-source exception, using a tree-sitter repo map.
- Two reference designs show what the missing discipline looks like in practice:
earendil-works/pi(harness-v2) and the Python harnesshelmsman. - The gap is not academic: a harness whose own record of an invocation is missing or non-authoritative is a concrete instance of this failure class.