CryptoComLearn has published a long-form guide outlining a 14-step, three-layer framework for building a self-improving agent system around Claude Fable 5. The piece centers on the combination of loops, dynamic workflows, routines, and memory, arguing that most users still treat Fable 5 like a one-off prompt model even though Anthropic designed it for long-running agent sessions.
How Fable 5 is positioned and priced
According to the article, Claude Fable 5 launched on June 9, 2026 as the first publicly available Mythos-tier model, a level Anthropic places above Opus. The cited launch materials say Fable 5 can run for days inside agent harnesses such as Claude Code or Claude Managed Agents, handling multi-stage planning, delegating to sub-agents, and checking its own work.
The pricing reflects that tier. The article lists $10 per million input tokens and $50 per million output tokens, while keeping a 90% input token discount through prompt caching. Availability is described across Claude API, AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry, and usage-based Enterprise plans.
Self-improving is not self-learning
A major distinction in the piece is the gap between self-learning and self-improving systems. Self-learning would mean a model updates its own weights, which the article says Fable 5 does not do. Self-improving, by contrast, means the system around the model keeps accumulating lessons in state files, skills, and evaluation rules so the next run starts from a better position. The model is stateless. The surrounding system is not.
The write-up frames this as a four-layer stack: primitives such as the model, sub-agents, worktrees, and tools at the base; orchestration with /goal, Outcomes, dynamic workflows, and Routines above that; then memory structures including state files, skills, and knowledge bases; and finally a self-improvement layer built from visual checks, eval loops, and rule distillation.
Three core primitives for compounding behavior
On implementation, the article treats /goal and Outcomes as two versions of the same goal-driven loop. /goal is presented as a better fit for measurable tasks running locally in Claude Code, while Outcomes is aimed at longer jobs in Anthropic-hosted environments. In both cases, an independent grader decides whether the work meets the target, and failure triggers another iteration.
The guide also stresses that an independent validator sub-agent often beats self-critique. Its case is structural rather than stylistic: a separate validator only sees the output and the grading criteria, not the original reasoning path, which makes it less likely to favor the producer’s own conclusion.
Dynamic workflows, described as having launched in Claude Code on May 28, 2026, let Claude generate a task-specific JavaScript harness using primitives such as agent(), parallel(), and pipeline(). The article highlights three patterns as especially useful for self-improving systems: fan-out and synthesize, adversarial verification, and loop until done.
Worktrees and Routines for long-running agents
Once multiple agents are working in parallel, file conflicts become a practical issue. The article presents git worktree as a core isolation mechanism, allowing producers, validators, and parallel structural experiments to operate in separate working directories tied to the same repository history. For systems meant to run across days, this is framed as basic infrastructure rather than an optional tool.
Routines, which the article says entered research preview on April 14, 2026, are described as a cloud execution layer that stores Claude Code settings including prompts, repositories, connectors, and permissions. These routines can be triggered by schedules, API calls, or GitHub events inside Anthropic-hosted infrastructure. Example use cases in the piece include rerunning eval suites every morning, launching investigations after CI failures, and writing new patterns back into a skill after a pull request is merged.
Memory is what makes the system compound
The final section places the heaviest weight on memory. Drawing on Anthropic’s Continual Learning Bench 1.0, the article describes a five-stage progression of agent memory: fail, investigate, verify, distill, and consult. In one SQL exploration task cited by the piece, Sonnet 4.6 typically stopped at recording failures, Opus 4.7 progressed into partial verification, and Fable 5 in its stronger runs verified 22 out of 30 questions, for 73% coverage.
The practical container for this process is a state file such as STATE.md. The guide recommends dividing it into verified facts, general rules, open failures, lessons learned, and the latest session state, then enforcing two habits: write to the file before ending a session, and read it before starting the next one. Project-level memory lives there, while cross-project procedural memory is stored in Skills. The article’s conclusion on this point is simple: if lessons remain trapped inside a chat session, the system does not actually get better over time.