Agent harness tier list sparks debate over compaction and vendor lock-in

A tier list of agent harnesses has sparked a chain of arguments in the community, centered on the fairness of custom compaction and whether labs are using harnesses and compaction to create vendor lock-in. Nothing is settled yet, but several prominent developers (Sentry's CEO, LangChain's CEO) have publicly voiced frustration with the trend toward a closed lab ecosystem, which is worth watching.

Confirmed

• Paul (pvncher) challenged a certain agent harness tier list, arguing that any harness doing custom compaction on OpenAI or Anthropic models should be heavily downgraded, and asked how third-party tools should be handled.
• Sentry founder zeeg believes things are already bad if labs turn compaction into a vendor lock-in tool—some vendors have even removed reasoning—and that labs should focus on making great models rather than pushing customers away.
• zeeg also complained that today's coding agent harnesses and APIs are vendor-specific, making vendor-neutral solutions extremely painful, and that he often has to modify harness defaults (tools, prompt rewriting, etc.).
• LangChain CEO Harrison Chase explicitly pushed back on the view that "compaction is already locked up by the big labs, so innovation should only happen in MCP/plugin layers on top of existing harnesses." His core argument: the labs' built-in coding harnesses are not necessarily the best in the industry, and developers will gradually shift to open-source models and build their own harnesses.
• In another discussion, Sentry CEO David Cramer (zeeg) debated an Anthropic engineer over whether coding harnesses should be self-built. zeeg argued that if models keep getting locked down and harnesses can't be self-built, developers will struggle to innovate—and problems like compaction remain unsolved.

Why it matters

• The debate touches the core tension in the coding agent ecosystem: should model capabilities (like compaction and reasoning) be exposed to third-party harnesses via open interfaces, or serve as a moat for the labs' own toolchains?
• If top labs keep tightening model interfaces, developers (especially infrastructure vendors like Sentry and LangChain) may accelerate the shift to open-source models and self-built harnesses, reshaping the tooling landscape.

2026-10-06 ~ 2026-10-06 · 11 related posts

• [source] Agent harness tier list sparks debate: custom compaction harnesses should be discounted — pvncher · 2026-10-06
• zeeg warns labs against vendor lock-in via compaction: 'focus on models, not locking customers' — zeeg · 2026-10-06
• [source] OpenAI ships compaction in Responses API, sparking vendor lock-in debate among developers — pvncher · 2026-10-06
• OpenAI ships compaction in Responses API; devs debate if it kills differentiation — zeeg · 2026-10-06
• Debate: Coding Agents Can't Compete When Core Harness Abilities Are Locked — pvncher · 2026-10-06
• [source] LangChain CEO: Lab coding harnesses aren't best; devs will move to open models to escape lock-in — zeeg · 2026-10-06
• Sentry CEO vs Anthropic engineer: should you build your own coding harness? — pvncher · 2026-10-06
• Dev argues agent harnesses are vendor-locked, making neutral AI coding tools a pain to build — zeeg · 2026-10-06
• Models are trained on native harnesses; generic abstractions lose performance — pvncher · 2026-10-06
• Counterpoint: wrapping native harnesses only helps generic IDE use cases — zeeg · 2026-10-06
• Debate: Do custom agent harnesses beat building on the Responses API? — pvncher · 2026-10-06