@FinanceYF5: 3/ Building the compound stack from bottom to top four layers. Bottom layer is primitives: Fable 5, sub-agents, worktree - most people only encounter this layer. Second layer is orchestration: goal loops, dynamic workflows, cloud Routines. Third layer is memory: state files, Skills, knowledge bases. Top layer is self-improvement: visual self-...

3/ 🏗 Build the compound stack from the bottom up with four layers

The bottom layer is primitives: Fable 5, sub-agents, worktree—most people only encounter this layer. The second layer is orchestration: goal loops, dynamic workflows, cloud-based Routines. The third layer is memory: state files, Skills, knowledge bases. The top layer is self-improvement: visual self-checks, evaluation loops, rule distillation. https://t.co/iTu1dGA3Gb

Regret not using Fable 5 before it was shut down? It ran continuously for 6 days

1/ Most people treat Fable 5 as a faster chat box

Someone let a Fable 5 agent run continuously for 6 days with no one at the helm, and only then wrote down the conclusion: 90% of people only use 10% of its capabilities. It was built to run for days, yet people use it for just a few minutes.

2/ The real dividing line: self-learning vs self-improvement

Self-learning is the model changing its own weights. Fable 5 does not do this, and currently no production model does. Self-improvement is the system outside the model compounding: each run writes lessons into memory, skills become sharper with use. The model stays the same, but the environment gets smarter with every run.

3/ Build the compound stack from the bottom up with four layers

The bottom layer is primitives: Fable 5, sub-agents, worktree—most people only encounter this layer. The second layer is orchestration: goal loops, dynamic workflows, cloud-based Routines. The third layer is memory: state files, Skills, knowledge bases. The top layer is self-improvement: visual self-checks, evaluation loops, rule distillation.

4/ Don’t throw everything at Fable 5 to run

Its cost per token is about 5 times that of Opus 4.8 ($10 per million input, $50 per million output). Let Fable 5 be the orchestrator, Sonnet 4.6 do the heavy lifting, Haiku 4.5 act as the scorer, and automatically fall back to Opus 4.8 when blocked by the safety classifier.

5/ Never let the model grade itself

Anthropic’s own experiments: the version with an independent validator dared to make bigger changes, pushing a failed experiment to maximum victory; the self-scoring version only dared to tweak one safety parameter and quit early. The agent that writes code should never be the one that grades it.

6/ The five stages of memory: Failure → Investigation → Verification → Distillation → Retrieval

Sonnet 4.6 mostly stops at the first step, piling up failure notes no one ever reads. Fable 5 can complete the entire process; at its peak, verification coverage reached over 70%, distilling facts into reusable rules. The gap is not in the model, but in whether you have state files.

7/ Self-improvement is a property of the system, not the model

In every experiment that proves this, the models on both sides are identical. What changes is the outer system: the verifier, state files, evaluation loops. Pick a layer you haven’t implemented yet, add it tomorrow, then add the next one.

Read the original article:

That’s all.

If you like this topic:

1. Follow me (@FinanceYF5)
2. Like and retweet the first post below