How to Build Self‑Correcting Loops with Claude Code’s Fable 5

Claude’s Mythos series models such as Fable 5 have changed how many Anthropic employees work. Two practical techniques are presented to fully exploit these models.

1. Self‑Correcting Loop

Iteratively optimizing a model based on evaluation metrics is a common way to improve task outcomes. In Claude Code, the /goal command and the Outcomes feature of Claude Managed Agents provide the core capabilities for this approach. By defining a clear goal or scoring rubric, Claude receives feedback after each execution, corrects its behavior, and repeats until the target is met.

During testing, a simple “Parameter Golf” challenge was used. The challenge requires training the best possible model within 10 minutes on eight H100 GPUs, keeping the model under 16 MB. This mirrors Karpathy’s auto‑research project, which tests an agent’s ability to edit a training script, launch training, read logs, obtain a score, and decide the next experiment.

Using a Claude Managed Agent (CMA) sandbox with eight H100 GPUs, Fable 5 was compared to Opus 4.7. CMA supplies scheduling and a hosted sandbox that fits Fable 5’s long‑running tasks. A crucial detail is the evaluation mechanism: when the model judges its own output, inconsistencies arise, as noted by Prithvi Rajasekaran in an Anthropic engineering blog.

To address this, a separate scoring sub‑agent was employed via Outcomes, allowing independent evaluation. Each test run supplied a scoring file with nine verification criteria (e.g., benchmark execution, completing 20 experiments). The task ran up to eight hours, and the Outcomes scorer only allowed termination after all criteria were satisfied.

Results showed Fable 5 improved the training workflow about six times more than Opus 4.7. When experiments were split into structural changes (e.g., model architecture) and numeric tweaks (e.g., constant parameters), Fable 5 favored large‑scale structural optimization and demonstrated greater resilience, such as overcoming quantization rollback issues.

Opus 4.7 achieved only modest gains in the first round and then repeatedly applied the same numeric‑adjustment pattern.

2. Memory Across Sessions

Fable 5’s second major advantage is its external, cross‑session memory. Claude can write memories during a session that are later retrieved in subsequent sessions.

Using the newly released Continual Learning Bench 1.0 (by @pgasawa), a task was selected that required an agent to connect to an SQL database, answer a series of independent questions, and invoke memory as needed. The benchmark compares Fable 5, Opus 4.7, and Sonnet 4.6.

All agents ran within a CMA that provided a shared mounted filesystem for memory access.

The required workflow for effective memory use is:

Failure (record the error)

Investigation (analyze the cause)

Verification (turn the diagnosis into a checkable fact)

Extraction (create a general rule)

Reuse (call the rule directly instead of re‑deriving)

Observations:

Sonnet 4.6 stopped after the first step, merely logging failures and unverified guesses, and rarely accessed history without custom memory prompts.

Opus 4.7 reached roughly the third step, annotating uncertainties (e.g., “price might be in cents? verify”), but its verification coverage ranged only 7 %–33 % (median ≈ 17 %).

Fable 5 completed the full workflow, achieving up to 73 % verification coverage (22 of 30 criteria) and extracting reusable rules that improved subsequent tasks.

Thus, designing a loop workflow—letting the model self‑correct via /goal or Outcomes and managing context through memory—outperforms ad‑hoc prompt engineering.

The author encourages readers to test Fable 5 on more demanding tasks, using loops for self‑correction and memory management, and points to the official Claude Code documentation and the built‑in /claude‑api skill for further guidance.