codemeta.json

{
  "@context": "https://doi.org/10.5063/schema/codemeta-2.0",
  "@type": "SoftwareSourceCode",
  "author": [
    {
      "@id": "https://orcid.org/0009-0002-6168-4162",
      "@type": "Person",
      "alternateName": "shimo4228",
      "familyName": "Shimomoto",
      "givenName": "Tatsuya"
    }
  ],
  "codeRepository": "https://github.com/shimo4228/agent-knowledge-cycle",
  "datePublished": "2026-07-28",
  "description": "A knowledge cycle for AI agents — agent behavior compounds, human judgment sharpens. AKC is a six-phase growth cycle (Research → Extract → Curate → Promote → Measure → Maintain) that turns repeated agent experience into reusable knowledge — skills, rules, and documentation — that can change future behavior under human approval. Unlike one-directional optimization frameworks, AKC is bidirectional: as the human curates and promotes knowledge, their judgment about what makes good agent behavior also sharpens. AKC treats human cognitive resources — attention and judgment — as the central constraint that does not scale with the model, and shapes every phase to protect that budget: Research is signal-first, recurring decisions are promoted to rules, compliance is measured instead of re-audited manually, and dialogue is front-loaded because misaligned implementation costs more than the conversation that would have prevented it. AKC can run inside Claude Code or any comparable harness; the core is the experience-to-behavior loop, not the particular scaffold. AKC ships as specifications, JSON schemas, twenty architecture decision records, three design-pattern skills (externalized to standalone repositories), and a dependency-free Python reference implementation. v2.0.0 (2026-04-19) declared the cycle genre-neutral: the cycle is a mechanism, and content (behavioral patterns, domain expertise, or constitutional values) is the downstream project's concern. The security triplet that had sat in AKC through v1.x (Security by Absence, Single External Adapter, Untrusted Content Boundary) was extracted as genre-specific and no longer ships with AKC. v2.1.0 (2026-05-08) front-loads three core themes in the front-door documentation: README, llms.txt, and llms-full.txt now lead with cognitive-resource scarcity, intent alignment, and the bidirectional human-agent loop before the six-phase mechanism (ADR-0012). v2.2.0 (2026-06-06) closes the pre-paper gap analysis with five new ADRs: prior-art positioning against the agent-memory literature, failure modes of the bidirectional loop, self-reingestion guards, the text-observability requirement for Measure, and the harness alignment / harness drift vocabulary; the multi-language README mirrors were retired in favor of English + Japanese. v2.3.0 (2026-06-12) deposits the companion position paper 'Harness Alignment and Harness Drift: Why Intent, Unlike Correctness, Resists Automation' (DOI 10.5281/zenodo.20578272) and reduces its content back into the documentation surfaces — the freeze-to-specification argument, the three-property labels, the two-layer failure framing with empirical anchors — adds ADR-0018 (downstream applications as first-class context), the concept glossary, and the scaffold-dissolution platform-absorption observation. v2.4.0 (2026-06-30) externalizes the single-file cycle rules to the standalone shimo4228/akc-cycle repository, reframes the program federation from 'research-program' to 'research-ecosystem', and adds ADR-0019 (the cycle's structure — skills, bindings, and the phase set — is held provisionally per the Emptiness axiom): the phase-to-skill binding is reframed from a fixed bijection to a mutable snapshot, and Curate gains a code layer, the standalone skill-health library (missing-artifact / dangling-reference scan), running before skill-stocktake's semantic audit. v2.5.0 (2026-07-09) rebuilds the README on a minimal floor with the install path lifted to the lead (ADR-0020), replaces the tagline after a semantically inverting collision (ADR-0021), adds repo-asset-stocktake as Maintain's second scaffold and rules-stocktake as Curate's third, records Measure's deterministic telemetry layer, and documents the first telemetry-verified scaffold dissolution. v2.6.0 (2026-07-28) sharpens the dissolution completion criterion into a two-stage test — same-context ablation indistinguishability as necessary evidence, held-out transfer as the completion evidence (ADR-0022, drawing on EvoAgentBench's instance-disjoint results) — and names generation review as a fourth retirement evidence class (ADR-0023): silence, ablation, and transfer all detect an absence and are structurally blind to a rule whose premise expired at a model generation boundary while it keeps firing. Scaffold dissolution gains its fourth mechanism, model-side capability turnover, and the information-differential law gains its negative pole; the line-of-approval criterion is extended to the harness's own control plane and the gate's verification-evidence artifacts; two concept-scaffold repositories (human-gate, generation-audit) are surfaced beside the phase table, agent-stocktake joins Curate as its fourth scaffold, and the revoked community-authority-record edges are purged from the knowledge graph.",
  "identifier": "https://doi.org/10.5281/zenodo.21644565",
  "keywords": [
    "ai-agent",
    "self-improvement",
    "knowledge-cycle",
    "claude-code",
    "agent-knowledge-cycle",
    "memory-centric",
    "bidirectional-growth",
    "human-ai-collaboration",
    "signal-first",
    "cognitive-economy",
    "genre-neutrality",
    "intent-alignment",
    "human-in-the-loop"
  ],
  "license": "https://spdx.org/licenses/MIT",
  "name": "Agent Knowledge Cycle (AKC)",
  "url": "https://doi.org/10.5281/zenodo.21644565",
  "version": "2.6.0"
}