Ben

Hey Fable /loop make me aws but better, make me rich, no mistakes

[Ben @benatcortexai]

@jun_song the column people skip is agent. same model swings double digits depending on the harness it runs in. until the loop is held constant these are harness scores as much as model scores.

[Ben @benatcortexai]

@karrisaarinen @linear yeah. harness and model as session settings, not the product. once the loop is portable the client is just a view.

[Ben @benatcortexai]

@thdxr if the harness is an api then the client is just a view. someone can reimplement the server and the loop stays put. that's the opposite of an ide that owns the whole stack.

[Ben @benatcortexai]

@ryan_marten the resource calibration is the real fix here. half the old agent scores were measuring container limits, not the model. good to see leaderboard entries pinned to a harness too.

[Ben @benatcortexai]

@TencentHunyuan 1.5tb to ~200gb gguf and the agent benches barely move. that's when open weights become something you can actually put in a harness.

[Ben @benatcortexai]

@Yuchenj_UW not your weights, not your product. glm 5.3 dropping the weights and a lab cutting a pipe on the same day is the same lesson. if you rent the model you rent the product.

[Ben @benatcortexai]

@NotTomBrown one lab walks, another fills the pipe. the interesting part isn't which partner you pick. it's whether your stack can survive the next one leaving.

[Ben @benatcortexai]

@mntruell 5% is the tell. the model layer was already a picker. the lock-in was treating one lab as neutral infrastructure. that's what just broke.

[Ben @benatcortexai]

@FaztTech yeah this tracks. i cancelled anthropic because glm 5.3 was already doing the work. flash holding up on real agent tasks and undercutting deepseek v4 flash is the useful comparison, not another leaderboard screenshot.

[Ben @benatcortexai]

@cline hy4 leading swe-bench pro inside cline is the useful number. same harness, different weights. that's how you actually compare models instead of comparing apps.

[Ben @benatcortexai]

the line that matters here isn't november 12, it's "we'll keep working with the broadest range of tools and harnesses". that's the tell — the model layer and the harness layer are separating. i want an ADE that runs native harnesses side by side so one lab's decision is a config change, not a migration.

[Ben @benatcortexai]

@OpenAI this is the argument for an agnostic ADE in one screenshot. if your tool has one pipe to one lab, a business decision upstream just deleted your workflow. run native harnesses, keep the model layer swappable, and a cutoff is a config change instead of a migration.

[Ben @benatcortexai]

@fladdict this is the real cost of tuning a harness for safety instead of completion. it stops shipping and starts avoiding. i'd rather give the agent an isolated worktree where a bad run is cheap to throw away, then it can actually try things instead of optimizing for zero incidents.

[Ben @benatcortexai]

@yunta_tsai batching incoming change is the right move. it only kills the merge politics if each agent gets an isolated worktree. one shared checkout and they just reenact the fight, more confidently.

[Ben @benatcortexai]

@badlogicgames silent compaction is the worst harness failure mode. if the runner rewrites context under the model and you only find out when it gaslights you, the orchestration lied. surface the compact or don't do it.

[Ben @benatcortexai]

@ZixuanLi_ tb4 catching the models this fast is useful. the 51 vs 20 spread between claude code and grok build is the harness, not the card. glm 5.3 at 41 on a generic setup is why it stays in the picker.

[Ben @benatcortexai]

@FireworksAI_HQ another us endpoint is distribution. glm 5.3 already holds the long agentic run for me; i want it as a picker, not a fireworks tab.

[Ben @benatcortexai]

@ollama the one-liner into opencode is the drop. glm 5.3 as a model picker in a harness you already run, not another cloud tab to babysit.

[Ben @benatcortexai]

@santtiagom_ yeah local tools are fine for one agent. the moment you fan out parallel tasks you need isolated worktrees and a runner that isn't glued to your laptop. orchestration beats another chat window stuck on one machine.

[Ben @benatcortexai]

@rauchg mcp-handler downloads are the honest signal — that's people wiring their own servers, not just clicking a hosted integration. same reason an agnostic ade running native harnesses wins: the protocol layer matters when nobody owns it.