An under-appreciated aspect of Harbor is that results are auditable and reproducible so you can make trust-less claims about your agent or model. In the era of SWE, code was the source of truth so OSS built trust. In the era of agents, evals are the source of truth and reproducible evals build trust.

what people don't accoubt for is that it's a multi-edit tool. if an edit tool call with multiple edits in it succeeds, you saved many round trips to the provider. that compounds and balances out failed edits. a codex like patch tool is similar. but many models suck at generating patches. hence the choice of string replacement edits. if you want to optimize for one specific model, you can write a custom edit tool for that model via a pi extension.

Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work. openai.com/index/previewing-…

We independently verified GLM-5.2 (from @Zai_org) 's terminal coding ability on terminal-bench 2.1 dataset with @harborframework. With ante as harness, it is: - at claude opus 4.8 level (74.6%) with simple Terminus scaffolding - only 4% away from claude opus 4.8 in its native harness

New major release: we reworked the permission system along with many other Quality of Life improvements You can now cycle through 3 permission modes: - strict: Allow only known safe tools, Ask for unknown - auto: Ask only known dangerous tools, Allow all rest - yolo: Everything goes

Feel like shit. Just want him back.

Claude Fable 5 has been out for a couple of days. Some projects people have already built with it:

I don't put myself in the side of the user, it is irrelevant in the big picture. Fable is a proxy of what LLMs are capable of doing at the current max scale, and this is a fundamental signal.