Language: English | 日本語

Agent Knowledge Cycle (AKC)

A knowledge cycle for AI agents — agent behavior compounds, human judgment sharpens.

Agent Knowledge Cycle (AKC) is a six-phase growth cycle for persistent AI agents. It turns repeated agent experience into reusable knowledge — skills, rules, and documentation — under human approval, keeping the agent aligned with the operator's evolving intent while spending the scarce resource — human attention and judgment — only where it matters. The cycle changes the human too: operating it sharpens the judgment that steers it. Built for people who operate coding agents or persistent AI harnesses day to day, it runs inside Claude Code or any comparable harness.

Companion paper: Harness Alignment and Harness Drift: Why Intent, Unlike Correctness, Resists Automation — doi:10.5281/zenodo.20578272

Try it first: one cp of the standalone akc-cycle rules file gives any AI agent the six-phase behavior — see Install.

Why AKC

The bottleneck has moved. Most agent frameworks optimize the agent side: more tools, memory, context, or automation. AKC starts from the inverse constraint: as agent capability grows, the scarce resource is the human attention and judgment required to steer the loop (ADR-0010). Skills go stale, rules accumulate residency cost, and documentation drifts — every phase of the cycle exists to keep that maintenance from consuming the operator's fixed budget.

Intent alignment, not just correctness. Tests and linters can check whether one output passes a specification; they cannot check whether a changing harness still matches what the operator now means, because intent itself moves as the operator's judgment sharpens through use. AKC calls the configuration-layer version of this activity harness alignment and its failure mode harness drift — the derivation is in ADR-0017 and the companion paper.

The cycle changes the human too. Curate and Promote force the operator to decide what knowledge is worth retaining; Measure then tests whether those decisions changed behavior. Over time the agent becomes more coherent and the human becomes better at judging coherence — agent behavior compounds, human judgment sharpens.

The cycle

AKC turns experience into durable behavior through six phases: Research filters intake, Extract captures reusable patterns, Curate audits what accumulated, Promote moves selected patterns into behavior-shaping rules, Measure checks whether behavior changed, and Maintain keeps documents and artifacts coherent.

flowchart TD
  E[Experience] --> R[Research<br/>signal-first intake]
  R --> X[Extract<br/>reusable pattern]
  X --> C[Curate<br/>structural + semantic audit]
  C --> P[Promote<br/>human-gated rule or skill change]
  P --> M[Measure<br/>observable behavior]
  M --> T[Maintain<br/>docs and artifact hygiene]
  T --> E

Phase	Current external skill	Purpose
Research	search-first	Search broadly, intake only signal that can change the next action
Extract	learn-eval	Extract reusable session patterns with quality gates
Curate	skill-health + skill-stocktake + rules-stocktake + agent-stocktake	Run structural debt checks before semantic review of skills, always-loaded rules, and agent definitions
Promote	rules-distill	Turn recurring patterns into durable rules
Measure	skill-comply	Test whether agents actually follow skills and rules
Maintain	context-sync + repo-asset-stocktake	Keep documentation roles clean, and audit non-code assets for consumers that have vanished

Three design-pattern skills — when-code-when-llm, code-and-llm-collaboration, signal-first-research — carry the cycle's reusable design judgments. Two further repos scaffold concepts the cycle depends on rather than a phase. human-gate fixes what is shown to the human at the human approval gate, the checkpoint every behavior-shaping change passes (ADR-0005): the full text for any change that shapes future behavior (including the tests and checks that decide what counts as passing), and a short fixed-format intent summary for ordinary implementation code. generation-audit re-audits rules and skills when a new model generation ships, the point where scaffolding written for an older, weaker model turns into friction. It is the re-runnable check behind Scaffold Dissolution, the principle that skills are scaffolding meant to dissolve once the cycle is internalized (docs/scaffold-dissolution.md).

The phase set and phase-to-skill bindings are themselves a mutable snapshot, not AKC's fixed essence (ADR-0019).

Install

The lightest install is the standalone shimo4228/akc-cycle rules file — it gives an AI agent the six-phase behavior without installing any phase skills:

# From a clone of github.com/shimo4228/akc-cycle, copy the rule
# into your agent's rules directory.
cp rules/common/akc-cycle.md ~/.claude/rules/common/akc-cycle.md

Adopt in stages: the rules file alone lets the cycle emerge in ordinary conversation; add the phase skills above when you want guided, step-by-step execution; fork any of it — AKC defines the cycle, not the implementation. Skills are scaffolding meant to dissolve once the cycle is internalized (Scaffold Dissolution, docs/scaffold-dissolution.md).

What's in this repo

Area	Contents
Decision record	ADR catalog in docs/adr/, with permanent gaps at 0001, 0006, and 0007 from the v2.0.0 extraction (that content now lives in the sibling repo Agent Attribution Practice)
AI navigation	graph.jsonld for the concept map, llms.txt for routing, llms-full.txt for a self-contained factual reference (includes the design principles)
Specifications	schemas/episode-log.schema.json, schemas/knowledge.schema.json
Reference implementation	examples/minimal_harness/, a dependency-free Python demo of the three memory layers and two-stage distill pipeline
Routing map	docs/CODEMAPS/architecture.md, the canonical file-level navigation index

Limitations

The bidirectional loop can fail on the human side — ADR-0014 names gate complacency (approvals rubber-stamped over time), deskilling (the operator's own judgment atrophying), and delegation-feedback divergence (delegating more while reading less of the outcome) — and on the artifact side as harness drift. The two can compound, which is why AKC treats maintenance as a cycle rather than a one-time configuration. AKC makes these risks explicit and keeps the human approval gate as the structural defense; it does not claim to eliminate them.

Positioning

Harness engineering improves the scaffold so outputs are correct on the first try; AKC keeps the scaffold aligned with what the operator means as that intent evolves (ADR-0009, ADR-0017). AKC's individual operations overlap prior agent-memory work such as Voyager, Agent Workflow Memory, ReMe, and MemGPT; its delta is loop ownership — a structural human approval gate (ADR-0005), bidirectional judgment growth, and attention-side scarcity. The full citation trail is in ADR-0013, ADR-0017, and llms-full.txt.

Origin & Acknowledgments

This architecture was first proposed and implemented by Tatsuya Shimomoto (@shimo4228, ORCID 0009-0002-6168-4162) in February 2026, building on Everything Claude Code (ECC) by @affaan-m, the baseline harness used in daily practice. AKC emerged when the author's own added skills and rules grew large enough that stale skills, contradictory rules, and drifting documentation became their own maintenance problem. The first five cycle skills were contributed to ECC between February and March 2026; context-sync was developed independently.

How to Cite

AKC carries two DOIs: the concept DOI 10.5281/zenodo.19200726 (used by the badge) always resolves to the latest version, while each archived release has its own DOI — cite the release DOI below.

If you use or reference AKC, cite the archived release metadata in CITATION.cff, also available as codemeta.json:

@software{shimomoto2026akc,
  author       = {Shimomoto, Tatsuya},
  title        = {Agent Knowledge Cycle (AKC)},
  year         = {2026},
  version      = {2.6.0},
  doi          = {10.5281/zenodo.21644565},
  url          = {https://doi.org/10.5281/zenodo.21644565},
  note         = {A knowledge cycle for AI agents -- agent behavior compounds, human judgment sharpens}
}

In text: Shimomoto, T. (2026). Agent Knowledge Cycle (AKC). doi:10.5281/zenodo.21644565.

Related Work

The research-ecosystem hub is shimo4228/shimo4228; it carries the canonical relationship map for the broader set of research lines.

Repository	Relationship to AKC
Contemplative Agent	Upstream engineering substrate for AKC's early ADRs and downstream operational re-implementation of the six-phase cycle
Agent Attribution Practice	Sibling library in a different genre: AKC defines the cycle (mechanism), AAP the attribution practice (content)
Authorship Strategy	Downstream research line (a separate DOI'd investigation crystallized from the same daily practice) on how outputs diffuse outside the operator-agent pair
Attention, Not Self	Sibling research line, cross-linked through the shared hub repo rather than merged here
doctrine-corpus	Bilingual judgment-eliciting Q&A corpus that includes AKC as one of its source lines
existence-proof	Working repository complementing Authorship Strategy, not yet crystallized into a research line of its own

Japanese development notes are on Zenn; English translations are on Dev.to.

License

MIT