What my feed was thinking about this week, plus my own notes.
Mainly two things this week: designing agent loops properly, and the whole Claude Fable situation.
Designing agent loops
Best way to think about it: draw the flowchart first. Before writing prompts or wiring up agents, figure out the deterministic flow you want the thing to follow.
Stuff you wanna pack into the loop as much as possible:
- Deterministic steps - so the system behaves predictably instead of pure vibes
- Verification steps - how the agent actually checks its own work
- Good tool calls - tests, logs, screenshots, repo inspection. Better signal in -> better output.
Basically:
goal → agent acts → verifier checks → state/memory updates → policy decides next action → repeat or stop or escalate
If you find yourself running the same multi-step workflow manually again and again, that’s your cue to turn it into a DAG. Claude Code’s dynamic workflows feature can generate a DAG on the fly, which is fine for quick exploratory stuff. But if you are putting something in prod, you gotta own the DAG yourself - the stages, checks, retries, stop conditions, review gates. Don’t outsource that part to the model.
Fable
Fable is legit insane. Tried it myself and it’s 100% worth spending 2 minutes looking at what it can do.
I have a few benchmark project ideas I throw at every new model. One of them is an interactive tool to teach spin in table tennis. Every frontier model before Fable completely fumbled the physics + 3D + UI together. Fable nailed it: https://srijan845.gumroad.com/l/spin-lab
Honestly if you didn’t feel a big jump with Fable, you’re probably not giving it a hard enough problem.
And then the drama happened. Days after release, the US government banned it. Apparently there was a jailbreak that Anthropic didn’t think was a big deal, USG disagreed, and shut it down. What a mess.
Also Fable was super expensive while it was live, so the pattern everyone settled on was:
- Use Fable for planning, architecture, taste/UI review, spatial reasoning
- Use GPT-5.5, DeepSeek, or Kimi as the worker to actually crank out the code
Other stuff from the week
- OpenRouter Fusion - OpenRouter exposed their Fusion feature as an API model. Basically a council-of-LLMs setup where multiple models contribute to the answer. They claim it matches Fable 5 solo.
- Google Open Knowledge Format - https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md - looks like the next version of what LLMWiki was doing. Structured reusable context for agents.
Also saved this image from somewhere on Twitter (forgot the source), good mental model for how much autonomy to give your models:
