Ghostty creator Mitchell Hashimoto details three-stage AI workflow using Fable xhigh and GPT 5.5

@mitchellh @mitchellh how do you share the plans between them?

@exempt_torch Very detailed, I prompt my planner (read-only Fable agent) for extra details including exact files and line numbers and so on.

@kfirgollan Copy and paste. Sub-agents are probably better but too lazy for this experiment lol.

@poojabnf I read the code

@icalvn Its not sophisticated. I have a "judge" OpenCode agent that has a system prompt that's basically "Look at the diff on this branch or unstaged set of changes and assess its quality. Look for any unhandled edge cases, obviously bad performance decisions" etc etc.

same, but i've been asking Fable just to orchestrate the GPT 5.5 agents using the claude.md in this thread from Theo:

It does a really good job of it. Been blasting out work today that was impossible with prior models (needed to implement various sub-components individually, being very strict on the agent; fable can just farm it out to GPT 5.5 in parallel, fast, and doesn't get slop)

I recently started experimenting with simple message passing between agents. Create some mechanism, say a tool, to read/write messages. I use it to have a planner write a plan then post a message, implementer read it, implement and post back. Makes connecting codex<->claude code essentially transparent and much better experience of using it. Also great enabler for cross harness loops

@mitchellh https://github.com/stablyai/orca seems like an interesting way of coordinating such agent orchestration workflows

@mitchellh Doesn't that end up costing just as much as just having Fable do the implementation itself? How much compression do you get by having it create such a detailed plan rather than just implementing it?

If it's just the cost of reading/writing files, hashline edits solve this!

@icalvn Oh also, this agent is read-only (has no access to file writing or shell tools). So it can only produce a report I can then circle back to GPT and so on. All manual atm.

@mitchellh Is GPT 5.5 a better coder than Opus 4.8? Or does it follow instructions better?

@mitchellh @kfirgollan You can use Claude code as a subagent of codex with Apple + J

You’ll probably find similar results with fable low or high tbh as coordinator

@mitchellh Same but I just tell Fable to use one Opus sub-agent at a time for implementation.

@kfirgollan @mitchellh We are almost about to invent unix sockets 🤣

@mitchellh How detailed are the plans?

@mitchellh I am currently trying fable-5 as planner/architect + opus-4.8 as coder.

I have the feeling that opus-4.8 writes better code than gpt-5.5

@mitchellh you can use codex plugin for cc to delegate task

https://github.com/openai/codex-plugin-cc

@mitchellh @kfirgollan Not sure which harnesses you're using, but you can run Codex in MCP mode with Claude Code. I've been experimenting with this since Fable was back and it's been really good at orchestrating Codex as sub-agents, even for large-ish scale refactoring

@mitchellh Are you actually getting any work out of Fable that you couldn't get with another model? I've seen zero real work done with it so far, just threejs games etc.

@mitchellh What exactly is xhigh - is there a token limit? Are we just gonna not acknowledge this.. then sonnet 5 xxxxhigh could also match fable for what it's worth