Anselme

@signulll GPT-Live is really good. The device will be next level.

unsurprising, the vision was to always bring her to life. i am very very curious how much they will lean on emotional companion aspect of this now that gpt live is really good. that will be a huge difference because apple would never lean into that (in fact they have explicitly said they would not), & google can never ever execute a product like this even if they wanted to. obviously alexa has a lot of users but that is going absolutely nowhere. those are the only hardware makers that are relevant in this space. the dynamics of this are even more interesting because jony ive went on a stripe stage & lamented how addictive the iphone is, imagine building an emotional companion that is even more addictive as his next act. that would be incredibly ironic.

but it also takes much, much longer to get things done connsidering to go back to 5.5

GPT 5.6 Sol seems to have significantly better judgement than 5.5

@zenorocha @resend The API should should have a built-iin agent that can execute natural language requests end to end. Essentially, which and how endpoints are chained becomes an implementation detail. The response simply returns proof of execution.

we're starting a new project internally... it's called *The Headless Dashboard Initiative* the goal is to make sure that everything you can do on the @resend dashboard, you can also do via: • API • SDK • MCP question: which endpoints should we build first?

@RhysSullivan i agree. though, the general consensus (misconception?) seems to be that MCP tools should be a subset of OpenAPI spec for an API, or just composite tools combining multiple endpoints... neither of which is lossless.

it's worth calling out for any companies wanting to do this, the best way is an OpenAPI spec that automatically gets converted into an SDK, MCP, and CLI easiest way to acheive this

given the complexity of the LLM-era software (security is a big part of it), we will actually see more open-source software, not less as many claim any non-trival use cases (agentic stuff) are way too hard for the average team

@andrewqu @zeeg Essentially, consistent, predictable behavior, when nothing has changed.

@zeeg Repeatability is the better word

The goal of using agents is determinism Skills are one, lightweight + quick way to add a little more determinism to your workflows

@theMissionBoy developers.cloudflare.com/changelog/post…

@mattzcarey docs or blog post link?

First class MCP elicitations support is in agents@latest! Both url and form modes :)

SBF: Semantic Backend Framework

after it mistakenly said the eval passed

gpt 5.6 medium seems to be sufficiently smart for most tasks, without draining your balance

@emollick many people seem unhappy with this adhoc communication style x.com/RhysSullivan/s…

Planning for how to use AI is a lot easier if there is some clarity about what to expect in the future. Even a "we fully intend to keep extending this week by week but may need to stop under the following conditions and here is what the current status is" would be better.

@0xblacklight Sol is especially good for model-facing code and prompts when you have evals, it's much better at diagnosing issues and fixing them correctly, including changing schema and prompt instructions 5.5 is mid at this

@0xblacklight Sol is also taking much longer for me but after ~3 hours without much supervision/steering, the results are better than what i usually get with 5.5. in terms of code design and overall quality

man y'all thought gpt-5.5 overthinks? just went in circles with 5.6 for like 2 hours bc it kept dispatching subagents for literally everything all the time plus it is super trigger-happy with skills switched back to gpt-5.5

gotta hope @thsottiaux will reset before monday did way too much with Sol this weekend and done for the week already on Pro plan :(

wat where do the training / rl data for these models come from

@zeeg same x.com/theMissionBoy/…

fixing some slop

i will switch to whichever coding harness best at enforcing "the easiest code to maintain is the code that was never written” for agents have to actively fight bloat like never before #coding