Why Coding Agents May Compete on Harnesses, Not Models

Summary

The article examines a shift in Coding Agent competition from model selection to Harness design: the surrounding system that manages context, tools, state, permissions, verification, and recovery. It argues that Claude Code, Codex, Cursor, and other products increasingly share a similar Agent Loop, while open-source and personal Harness projects are making the basic components widely accessible. Several practitioners say everyday coding tasks may no longer expose large differences between current models, making the way components are combined more important than the model alone. TiDB’s approach is described as a thin Agent Loop and a thick control plane, with persistent workspaces, sandboxing, permissions, checkpoints, rollback, quotas, and multi-tenant isolation built around an open-source core. Its Cloud Filesystem reportedly reached production in three months and has supported more than several million Agent Workspaces, with agents handling the implementation and review described in the article. Tencent Research Institute’s view emphasizes knowledge engineering: a Harness must understand existing systems, expose project constraints, support repeated reflection, and connect code generation to deterministic compilation and testing. The article is skeptical of making multi-agent orchestration the default. Communication and state synchronization can move complexity rather than remove it, so multiple agents are most justified for context pressure, parallel exploration, or cross-checking. The participants instead identify long-horizon stability as a major differentiator: systems must detect errors early, limit their propagation, and recover from a trusted checkpoint. Pi’s planned Harness v2 is cited as an example of persisting operation intent, tool state, and message queues for recovery after crashes. Floatboat reports that Harness design contributed 23% on long-horizon tasks with the model held constant, while advocating controlled self-improvement with isolation, evaluation, staged rollout, observability, and rollback. Finally, the article describes Coding Harness capabilities moving into office products. Coding offers clearer verification through tests and builds, whereas general work requires managing permissions, context, judgment, and social complexity. The shared foundation is goal understanding, tool selection, multi-step execution, state maintenance, and result verification.