Why Prompting Isn’t Enough: Designing Loops with Claude Fable 5

Lance Martin’s article, originally titled “Designing loops with Fable 5,” reveals that the real breakthrough in long‑task agents is not a smarter model but the loops that let the model receive external feedback, self‑correct, and reuse experience across sessions.

Self‑Correction Loops: Let the Environment Give Feedback

Martin contrasts ordinary prompting, which is a single request, with “loops” that embed a definition of “done” into the environment. In Claude Managed Agents, the /goal entry point and the “Outcomes” mechanism let the system tell the agent what success looks like and iterate until that condition is met. Anthropic’s Outcomes documentation states: “Tell the agent what ‘done’ looks like, and let it iterate until it gets there.” This external rubric‑driven feedback avoids the pitfalls of self‑critique, where a model’s own narrative can bias its assessment.

He cites the Anthropic Cookbook, noting that agents can produce polished‑looking outputs that still miss critical evidence, so an independent grader sub‑agent checks the work against a rubric, producing evidence such as opening cited URLs, matching quotes, and justifying failures.

Parameter Golf: Testing Fable 5’s Ability to Run Research‑Style Loops

Using OpenAI’s Model Craft Challenge (Parameter Golf), participants must train a model under strict constraints (≤16 MB artifact, ≤10 min training, 8×H100 budget) and evaluate on bits‑per‑byte on the FineWeb validation set. Martin runs this experiment with Claude Managed Agents, a sandboxed 8×H100 environment, and a rubric that requires nine conditions to be satisfied before the agent can stop.

The loop proceeds through steps: modify train_gpt.py, start training, poll logs, read scores, decide the next experiment, and adjust architecture, constants, quantization, or training strategy. He classifies experiments as “structural” (changing architecture) or “scalar” (tuning a constant). Fable 5 improves the training pipeline by roughly 6× compared to Opus 4.7 and prefers bold structural changes, even persisting through quantization regressions to capture larger gains.

Memory as an Outer Loop Across Sessions

Martin defines memory as an outer loop that persists beyond a single session. Using the Continual Learning Bench 1.0, he shows that effective memory follows five steps: fail, investigate, verify, distill, and consult. Instead of merely logging failures, the system turns them into verified facts and reusable rules that are consulted in future sessions.

The benchmark evaluates sequential tasks where performance is measured by a “gain” metric (reward minus a stateless baseline). In a SQL‑database task, the agent gradually builds knowledge of schemas, units, and pitfalls, demonstrating that memory can transform a stateless model into a learning system.

Comparing Sonnet, Opus, and Fable

Martin compares three Anthropic models:

Sonnet 4.6 stops at the first step, storing only failure notes and guesses (e.g., “maybe prc instead of prc_usd?”) and rarely reuses them.

Opus 4.7 reaches the third step, creating schema references and flagging uncertainties (e.g., “possibly prc in cents? Verify.”) but its verification coverage is low (7–33%, median ~17%).

Fable 5 reaches the full loop, achieving verification coverage up to 73% (22 of 30 checks) and distilling learnings into reusable rules.

This contrast highlights that true online learning requires admitting errors, investigating causes, verifying hypotheses, distilling rules, and re‑using them—capabilities that Fable 5 exhibits more fully.

Official Release Context and Safety Constraints

Anthropic’s June 9 2026 announcement positions Fable 5 as a Mythos‑class model with safety safeguards, while Mythos 5 offers higher‑risk access. Fable 5’s pricing (US$10 per M input tokens, US$50 per M output tokens) underscores that long‑running loops consume significant resources.

Safety mechanisms include a fallback to Claude Opus 4.8 for high‑risk domains (cybersecurity, biology, etc.), triggered in less than 5% of sessions.

Implications for Agent Product Design

Martin argues that the decisive factor for reliable agents is system design, not model size alone. He proposes a checklist for evaluating agent products, including clear “done” definitions, concrete rubrics, independent graders, real environmental feedback, and robust memory loops.

He concludes that the future distinction between agent products will resemble the gap between operating systems and workflow engines: the ability to define goals, verify outcomes, retain and apply experience, and know when to stop.

Prompt starts the model; loop makes the model trustworthy.