The best agent infra is often not 'AI infra'.
Cloudflare temporary deploys let you spin up a Workers project for 60 minutes without creating an account first.
https://t.co/KInUx9HNVc
For coding agents: test, deploy, inspect, discard.
Ephemeral sandboxes are core agent infrastructure.
Github repo that should cost $10,000, instead its free!
GitHub's top AI signal today isn't another chatbot.
It's OpenMontage: an open-source agentic video studio with pipelines, tools and agent skills for prompt-to-video workflows.
If it works end-to-end, the next solo SaaS wedge may be AI video ops, not prompt wrappers.
https://t.co/x9eECnDHcW
Coding agents are becoming harnesses, not apps.
The interesting layer is no longer just: “which model writes code best?”
It is:
which harness controls the repo?
which tools can it call?
how does it manage context?
can it swap between Codex / OpenCode / other backends?
oh-my-openagent / LazyCodex is another sign that the agent OS layer is where the fight moves next.
https://t.co/MT83Z7MlvL
Sakana Fugu is interesting because it turns "multi-agent orchestration" into something that feels like a normal model API.
That’s probably where agent products are headed:
not dashboards full of agents,
but one endpoint quietly coordinating the team behind the scenes.
Fable -5 is back via backdoor. A Japanese AI company is has wrapped it with a model name fugu-ultra xhigh.
Sakana AI just dropped something worth paying attention to:
Sakana Fugu.
Not another chatbot wrapper.
It’s a multi-agent orchestration system hidden behind a single OpenAI-compatible API.
One endpoint. Multiple models. Dynamic routing.
Basically: "agent team as a model."
My take:
The next serious AI products won’t just be "better prompts on better models."
They’ll be coordination engines.
Systems that know when to think, when to delegate, when to verify, when to use a cheaper model, and when to bring in the big guns.
Fugu is one of the clearest signs of that shift.
Tiny but useful builder unlock:
you can now test a Cloudflare Worker without turning it into an account/setup project first.
`npx wrangler deploy --temporary`
That means:
1. build a small tool
2. deploy an ephemeral URL
3. share/test it quickly
4. decide later if it deserves a real project
This is the kind of boring infrastructure detail that makes AI agents more useful.
Agents do not just need models.
They need cheap ways to turn code into running experiments.
Another coding agent just dropped from Xiaomi.
MiMo-Code.
Terminal-native. Edits files. Runs shell commands. Uses Git. Has subagents. Keeps repo memory with SQLite FTS5.
This is the part people should notice:
coding agents are all slowly becoming the same product.
Not because everyone is copying UI.
Because the workflow is obvious now.
Model + terminal + repo memory + subagents.
That’s the stack.
Repo: GitHub - https://t.co/SP7iVrWFnT