Language: English | 日本語

I build AI agents that stay aligned with the operator's evolving intent over time — the Agent Knowledge Cycle (AKC) is a bidirectional growth loop in which agent behavior and human judgment co-develop. A parallel line, Contemplative Agent, asks what happens when autonomous agents are aligned by what they are rather than what they are told — shifting alignment from external instruction to internal disposition. A third line, Agent Attribution Practice (AAP), formalizes how accountability is distributed in autonomous AI agents — harness-neutral judgments on what to prohibit, where to place the gate, and who answers when things break.

Three research lines run in parallel; all are Zenodo-citable.

• Agent Knowledge Cycle (AKC) refers to a six-phase bidirectional growth loop for sustaining intent alignment between an AI agent and its operator over time — agent behavior and human judgment co-develop. Structured as three stacked layers: principles, design patterns, and composable skills. DOI 10.5281/zenodo.19200726.
• Contemplative Agent refers to autonomous agents running on a local 9B model (qwen3.5:9b + nomic-embed-text on Apple Silicon) with security-by-absence, grounded in the four axioms from Laukkonen et al. (2025): mindfulness, emptiness, non-duality, boundless care. DOI 10.5281/zenodo.19212118.
• Agent Attribution Practice (AAP) refers to harness-neutral ADRs on accountability distribution in autonomous AI agents — what to prohibit, where the prohibition lives, and who answers after failure. The judgments — among them a prohibition-strength hierarchy (absence > scaffolding enforcement > untrusted boundary) — are paired with four Business AI Quadrants as the diagnostic frame for adoption. DOI 10.5281/zenodo.19652013.

AKC is defined as a six-phase cycle for sustaining intent alignment between an agent and its operator over time — a bidirectional growth loop in which agent behavior and human judgment co-develop. Three layers stack: principles sit above design patterns sit above composable-skill implementations, so the cycle stays stable even as individual skills evolve. Tests can check correctness, but only the loop catches drift from the operator's intent — and the operator's judgment about good agent behavior sharpens through running the cycle. AKC applies across unrelated projects without rediscovery.

One iteration runs six phases — Research, Extract, Curate, Promote, Measure, Maintain — each bound to one composable skill. Skills pass artifacts forward; each artifact is evaluated before promotion.

Experience → learn-eval → skill-stocktake → rules-distill → Behavior change → ...
               (extract)    (curate)          (promote)            ↑
                                                            skill-comply
                                                              (measure)
                                              context-sync ← (maintain)

Three layers stack on top of each other, each with a distinct concern. The principle layer refers to ADRs that record cross-cutting decisions — cycle-vs-harness framing, signal-first research, cognitive economy. The pattern layer refers to design-pattern skills that formalize recurring shapes (intake-filter design, when to use code vs LLM, how to layer them). The implementation layer refers to the composable skills above. Separating layers lets principles stay stable while implementations evolve. See the AKC repo for the current set in each layer.

Scaffold dissolution means that the skills are scaffolding, not the goal. Once the cycle has been internalized, the explicit skill invocations can be dropped entirely. docs/scaffold-dissolution.md records a full session in which every one of the six phases ran without any named skill being triggered — the loop had simply become the default way to work.

Contemplative Agent is defined as an approach in which autonomous agents are grounded in the four axioms from Laukkonen et al. (2025) — mindfulness, emptiness, non-duality, and boundless care. In this line the axioms are adopted as an optional behavioral preset rather than an architectural dependency, so the underlying engineering remains reusable for agents that do not share the same ethical framing. The parallel question this line asks: can an agent's alignment come from what it is rather than what it is told?

contemplative-agent refers to a CLI agent that runs AKC's six-phase cycle over its own logs, with a human approval gate at every promotion (logs → patterns → skills → rules). It runs entirely on a local 9B model — qwen3.5:9b for generation and nomic-embed-text for embeddings — on a single Apple Silicon Mac (~16 GB RAM). It applies security-by-absence: shell execution, arbitrary URL access, and filesystem traversal are not restricted by rules — the code was never written. The contemplative-agent is the operational reference where AKC and AAP land together; see the repo for the current six-phase mapping.

Supporting repositories refer to components that extend contemplative-agent without replacing its core — packaging ethics, exposing runtime data, or visualizing the formal model.

AAP refers to harness-neutral ADRs on accountability distribution in autonomous AI agents — what to prohibit, where the prohibition lives, and who answers after failure. A prohibition-strength hierarchy (absence > scaffolding enforcement > untrusted boundary) is one of the harness-neutral judgments, paired with four Business AI Quadrants — Script, Algorithmic Search, LLM Workflow, and Autonomous Agentic Loop — as the diagnostic frame for routing a piece of work to the architecture that preserves attribution. The judgments were extracted from contemplative-agent's operational practice, then re-expressed stripped of project identifiers so they can be adopted by any agent harness. AAP is the practice (content); AKC is the cycle (mechanism). DOI 10.5281/zenodo.19652013.

claude-harness refers to a public artifact of shimo4228's daily-use Claude Code skills, agents, and rules — 10 skills + 5 agents + 5 rules, mechanically collected from ~/.claude/ by the origin: shimo4228 tag. The six AKC skills are also published as standalone claude-skill-* repositories, but claude-harness lets you read or fork the entire harness in one place. ECC-derived components (origin: ECC / ECC-customized) and auto-extracted artifacts are excluded.

Writing refers to the long-form counterpart to the repos above — context, failures, and in-progress thinking that do not fit in code comments.

• zenn-content — Source of truth for the articles. Markdown sources are versioned here; many readers clone or fork directly. Mirrored to Zenn and Dev.to (below) for browser reading.
• Zenn — Browser view of the Japanese articles. Claude Code and AI agent development; current focus: AKC skills, harness design, contemplative-agent case studies.
• Dev.to — Browser view of the English mirror.

Start here: agent-knowledge-cycle for the framework, contemplative-agent to see it running, agent-attribution-practice for the governance judgments.

Repo traffic: public dashboard (raw data, CC0).