michelle

@this_maru a week where we launch things things we're working on!

@michellechen what’s innovation week? don’t tell me all that was released during agents week was made within one week

if u were wondering what an innovation week feels like at cloudflare its like caramelldansen nightcore 4x speed

@this_maru shy >.<

@michellechen why no cf badge tho

now im literally her

@dillon_mulroy in genz we say motherrrrrr

@michellechen matriarch of ai

she is her

@theCTO i should have gotten theCEO just to mog u

@michellechen oh it’s youu!

thank you @yagiznizipli 🖤

An AI agent can write your code in minutes. But someone still has to review, merge, deploy, and monitor it. What if the agent could do that too? Feature flags are the missing piece. They let an agent ship code behind a flag, test it on real traffic, ramp the rollout, and kill it instantly if things break. No human in the loop until you choose to be. Today we're shipping Flagship to make this possible - feature flags native to @Cloudflare's network, OpenFeature standard. Move fast, break nothing. blog.cloudflare.com/flagship

we’ve just released memory for agents primitives in Agents SDK when you need them, managed when you don’t

that's my quant Unweight: a lossless compression system that can make model weights up to 15–22% smaller while preserving bit-exact outputs, without relying on any special hardware blog.cloudflare.com/unweight-tenso…

@cowchimp oo good point, looking

@michellechen Hey can we get parameter count \ model size in the model details table of a single model in the cf docs please? Sometimes it's mentioned in the model name or description text but it's not standarized. Could be even better if it's also in the filters in the main page 🙏

@tanny2109 openai is doing this with websockets developers.openai.com/api/docs/guide…

@michellechen Naive question but can’t agents just first load up tools, mcp etc just once and then delegate user queries to subagents or something to get full context all the time at the cost of latency and orchestration overhead which I think is feasible now

serving models is hard. serving extra large language models with good quality, throughput, reliability, pricing, and gpu utilization is really hard. a look behind the scenes of how we do it on workers ai this agents week. there’s always more to do, but we’re just getting started blog.cloudflare.com/high-performan…

@vikhyatk maladaptive daydreaming on bat and zillow every day

just gave a talk about bringatrailer dot com. many are saying it was the best talk they’ve seen in a decade

@yagiznizipli WOW cf alum network goes so hard. tysm!

@michellechen I’ll take a look

im literally her

@BraydenWilmoth questioning my existence

@michellechen but are you?

@irvinebroque im too embarrassed to show u

@michellechen I don't know what that means — but absolutely yes

@tanny2109 they prefix cache it! but yes it gets sent on every turn

I’ve observed my Hermes and openclaw agent sending a lot of stuff correctly pointed out here like tools, mcp server etc each time they’re talking to my local model endpoint and I thought that in case of model apis from OpenAI and Anthropic they must store all this tool and mcp info on servers to prevent this, I didn’t know this happens there as well.

@superhighfives @replicate @Cloudflare @corywilkerson goat

If you've wondered what I've been up to for the last couple months, wonder no more. blog.cloudflare.com/ai-platform/ Amped for @replicate to be shipping with our new mates, and grateful to @cloudflare for encouraging us to go for it. As @corywilkerson said, this is just the beginning—we're onboarding tons of models, making it easier to bring your own, and turning AI Gateway into a whole control and data plane. Best team, best week. Happy days!

and one more thing — claude opus 4.7 now live on ai gateway: developers.cloudflare.com/ai/models/anth…