Post

(I ask this as a Django fanboy with the utmost respect for you)

I find that comforting 😂 appreciate all your detailed blogging

Can confirm. Fable dispatching GPT 5.5 implementers is amazing.

How are you deciding when to trust the model's own token-budget choices vs. overriding it? I keep wondering whether that delegated frugality holds up once tasks get ambiguous about what 'lower power' actually means.

Interesting. I had a "chat" with Fable about token usage and this is what I came up with. A "scout" agent for grunt work that uses Sonnet. gist.github.com/miklb/bf2d9e...

I've been doing this with Opus for a while - it definitely saves tokens farming out to Sonnet.

I’ve been doing this to great effect. It’s using Fable very little and getting the same results. A side effect is that now I can continue the chat with the main agent while the subagents work. Similar to a team, but without the startup and shutdown costs.

that's how I started but still blew through 1/2 my session limit in one prompt. (I'm on the $20/mo plan).

Fable doesn't, I guess I need to "trust it" to verify that those weaker models it farms works out to aren't doing that though!

used this on the repository (so it got no new violations), and used the new extended checker only on the current file, i found out when i saw it using two tools and forbade that and removed the extended checker, it just refused to fix the resulting violations, so i changed my checker to trick it>

Yes, many times they do what a human would do. What the have learn from the training data... That's why plans are important, you can catch early the wrong design and correct it before the implementation starts. But small models will cheat anyways 😅

The trick is making the budget decision reversible. I’m happy letting the main agent delegate boring reconnaissance to cheaper subagents, but I still want a crisp handoff: what it checked, what confidence it has, and when it escalated back instead of guessing cheaply.

Good rule to break the default of Fable of using fable for subagents. But Sonnet is so token hungry now... I would say it is better to use Opus most of the times. Haiku needs an update, deepseekv4flash is so much more capable...

This feels like an important shift: managing AI by defining decision rights rather than prescribing every step. In business processes, the real design challenge may become deciding what the AI can determine independently, what requires validation, and what must always remain a human decision.

I use Hermes Agent with GLM 5.2 and other free models on the Nvidia Direct API for all of my web scraping, so I don't cut into my claude credits.

I was doing that with opus already and it would sometimes deign to trust sonnet though almost never haiku. Fable is WAY better at making these choices.

letting the model use its own judgement is probably the bit people underprice. for indie stuff, even a boring 20% token saving adds up fast when every experiment is paid for out of the same card