RealPGRESEARCH CHRONICLE · 読むための記録
分かれたAIの仕事を、責任が途切れない形で続ける
観測したこと
いまの解釈
残る不明点
次の確認
何が起きたか
一つの会話に仕事を集めるだけでは、専門ごとの責任や引き継ぎが曖昧になりました。そこで、役割ごとに誰が何を引き受け、次の担当へ何を渡すかを明らかにする形を試しました。
どう考えているか
分業で必要なのは会話を増やすことではなく、責任の境目と仕事の続き方を分かるようにすることです。RealPGでは、この役割ごとの作業場所を Lane と呼びます。
まだ分からないこと
この分け方が、他の会社やより広い種類の仕事にも役立つかは、まだ分かりません。
次に確かめること
担当が替わっても、何度も人がつなぎ直さずに責任を保てるかを観察します。
なぜ重要か
複数のAIが関わっても、仕事が宙に浮かず、誰が次を引き受けるかを確かめられるためです。
技術的な記録と証拠 / Technical provenance and evidence references
RESEARCH CHRONICLE · 2026-08-16–2026-08-25 · OPEN
1つのAIから、Lane-localな会社基盤へ
この記録が追っているのは、専門化したAI workを、一つの巨大な会話workerではなく、責任・authority・継続性を持つowner-local Laneへ分けると何が変わるか。
出発点では、仕事は一つの未分化な会話workerに寄っていた。 そこへ「専門化した仕事は、責任と文脈の継続を必要とした。」という圧力が加わった。ここでは後から整った制度を過去へ遡って完成扱いせず、当時から残るevidenceと現在のcanonical recordを分けて読む。
現在の解釈では、Laneの価値は『AIを複数に増やしたこと』ではなく、誰が何を変えられるか、誰が意味を決めるか、どこでhandoffするかを明示できることにある。 ただし、このLane構造が他企業・他domainでも同じ粒度で有効か。 この未解決点を残したまま、次のverificationへ進む。
研究課題
専門化したAI workを、一つの巨大な会話workerではなく、責任・authority・継続性を持つowner-local Laneへ分けると何が変わるか。
以前と圧力
以前はどうだったか
仕事は一つの未分化な会話workerに寄っていた。
何が足りなくなったか
専門化した仕事は、責任と文脈の継続を必要とした。
観察したこと
2026-08-16から25日にowner-local authorityが段階的に現れた。
- 2026-08-16以降、Machine、Mirage、Application、Nagiなどのowner-local Laneが段階的に分離され、repository mutation、semantic coordination、Product runtimeなどの責任境界が明示された。
- Application v6では、Application-origin Change IntentがMachineを直接targetできてしまうauthority gapが露出し、invocation-lane→target-lane bindingの必要性が観測された。
- Machine v13では、organization-wide semantic visibilityとrepository mutation authorityは別概念として扱うべきだというlearningがdurable化された。
- Lane-local validationが他Laneのsource treeへ依存すると、owner boundaryを越えて壊れる例もMirageで観測された。
そこから推論したこと
Laneは自律性を証明せず、責任境界を明示する。
- Laneの価値は『AIを複数に増やしたこと』ではなく、誰が何を変えられるか、誰が意味を決めるか、どこでhandoffするかを明示できることにある。
- semantic participant identityとmutation authorityを分離すると、Iris/Experimentのような非repository actorもCompany stateへ参加できる。
- owner-local boundaryは自律性の証明ではなく、責任の局所化とfail-closed routingの基盤である。
どう進んだか
- PRESSUREundifferentiated worker responsibility
context:20260912-050541:task.laneIndex - BOUNDARY_FAILUREorigin lane vs mutation target
application:v6:application-origin-target-enforcement-rendezvous:20260829 - INSTITUTIONALIZATIONsemantic participation vs mutation authority
machine:v13:semantic-participants-separate-from-mutation-authority-v0:20260901
変えたこと
- 開発責任をowner-local laneへ分けた。
- 開発責任をowner-local Laneへ分けた。
- cross-authority workを直接mutationではなくtyped coordination/rendezvousへ寄せた。
- semantic participantとrepository authorityを別概念として扱った。
その後どう変わったか
- 継続する専門作業の境界が得られた。
変えなかったこと
- Laneを独立した『会社』として扱っていない。
- 全Laneに同じmutation authorityを与えていない。
まだ分からないこと
まだ検証できていない: 一般企業への適用可能性。
- このLane構造が他企業・他domainでも同じ粒度で有効か。
- Lane数が増えた時のcoordination overheadが、単一workerより低く保てる範囲。
- Objective Lineage導入後、cross-Lane handoffでHuman reconnectionがどれだけ減るか。
Counter / 注意点
- Laneが増えること自体は価値ではなく、境界が過剰ならcoordination costを増やす。
- owner boundaryを守っても、parent ObjectiveのLineageが無ければHumanがLane間の因果線を保持する必要が残る。
次に確かめること
引き継ぎが繰り返しのつなぎ作業なしに権限を保つか観察する。
- 同じObjectiveが複数Laneを横切るepisodeで、handoff回数・Human wake数・authority violationを測る。
- Lineage導入前後でcross-Lane resume costを比較する。
なぜ重要か
分業を会話量ではなく責任で扱える。
- 専門化をprompt personaではなく、責任・authority・state continuityとして表現できるようになった。
- 誤ったtargetへの変更や他Lane sourceへの隠れた依存を、境界failureとして検出できるようになった。
このstatusは成功宣言ではなく、canonical recordの現在位置を示す。
証拠とprovenance
context:20260912-050541:task.laneIndexapplication:v6:application-origin-target-enforcement-rendezvous:20260829machine:v13:semantic-participants-separate-from-mutation-authority-v0:20260901mirage:v13:lane-local-validation-must-not-require-application-source:v0:20260904context:20260913-201241:public/development/index.jsongit:branch:lane/machine/20260816t021523z-e5d20dfd
Keeping responsibility continuous when AI roles are split
RESEARCH CHRONICLE · A record for reading
Observed
Interpretation
Unknown
Next check
What happened
Putting all work into one conversation made specialist responsibilities and handoffs unclear. This work tried a clearer arrangement: each role has a bounded job, and the next person can see what needs to be carried forward.
What we think it means
Division of work needs more than additional conversation. It needs visible responsibility and continuity. RealPG calls these role-specific work areas Lanes.
What remains unknown
It remains unknown whether this arrangement helps outside RealPG or across a wider range of work.
What we will check next
We will observe whether handoffs preserve responsibility without people repeatedly stitching the work back together.
Why it matters
When several AIs take part, work should not become ownerless or lose its way between roles.
Technical provenance and evidence references
RESEARCH CHRONICLE · 2026-08-16–2026-08-25 · OPEN
From one AI workflow toward lane-local company ownership
This episode asks: What changes when specialized AI work is divided into owner-local lanes with bounded responsibility and authority rather than one undifferentiated conversational worker?
The starting point was that work was concentrated in one undifferentiated conversational worker. The pressure became explicit: Specialized work needed continuity of responsibility and context. The article keeps later institutional interpretation separate from evidence that actually remains from the period.
The current interpretation is that the value of lanes is not simply “having more AIs”; it is making explicit who may change what, who decides meaning, and where handoff occurs. But whether the same lane granularity works in other companies or domains. That uncertainty remains part of the record rather than being edited out as the work becomes more readable.
Research question
What changes when specialized AI work is divided into owner-local lanes with bounded responsibility and authority rather than one undifferentiated conversational worker?
Before / pressure
Prior state
Work was concentrated in one undifferentiated conversational worker.
Failure or pressure
Specialized work needed continuity of responsibility and context.
Observations / What we observed
Owner-local authorities emerged across 2026-08-16..25.
- From 2026-08-16 onward, owner-local lanes such as Machine, Mirage, Application, and Nagi were separated gradually, making responsibility boundaries for repository mutation, semantic coordination, and Product runtime explicit.
- Application v6 exposed an authority gap in which an Application-origin Change Intent could target Machine directly, showing the need to bind invocation lane to target lane.
- Machine v13 made durable the learning that organization-wide semantic visibility and repository mutation authority are different concepts.
- Mirage also observed that lane-local validation breaks an ownership boundary when it depends directly on another lane’s source tree.
Hypotheses / Inferences
Lanes do not prove autonomy; they make ownership boundaries explicit.
- The value of lanes is not simply “having more AIs”; it is making explicit who may change what, who decides meaning, and where handoff occurs.
- Separating semantic participant identity from mutation authority lets non-repository actors such as Iris or Experiment participate in Company state without acquiring repository authority.
- An owner-local boundary is not evidence of autonomy; it is infrastructure for local responsibility and fail-closed routing.
How it unfolded
- PRESSUREundifferentiated worker responsibility
context:20260912-050541:task.laneIndex - BOUNDARY_FAILUREorigin lane vs mutation target
application:v6:application-origin-target-enforcement-rendezvous:20260829 - INSTITUTIONALIZATIONsemantic participation vs mutation authority
machine:v13:semantic-participants-separate-from-mutation-authority-v0:20260901
Interventions / What changed
- Separated development responsibility into owner-local lanes.
- Development responsibility was divided into owner-local lanes.
- Cross-authority work moved away from direct mutation toward typed coordination and rendezvous.
- Semantic participation and repository authority became separate concepts.
Outcomes
- A boundary for continuing specialized work was established.
What did not change
- A lane was not treated as an independent “company.”
- All lanes were not given the same mutation authority.
Unknowns / What remains unknown
Not yet verified: applicability to general companies.
- Whether the same lane granularity works in other companies or domains.
- How far lane count can grow before coordination overhead becomes worse than a single worker.
- How much Objective Lineage reduces Human reconnection during cross-lane handoff.
Counter-evidence / cautions
- More lanes are not valuable by themselves; excessive boundaries can increase coordination cost.
- Even with correct owner boundaries, a missing parent-Objective lineage can leave Humans carrying the causal thread across lanes.
Next verification
Observe whether handoffs preserve authority without repeated stitching.
- In episodes where one Objective crosses multiple lanes, measure handoff count, Human wakes, and authority violations.
- Compare cross-lane resume cost before and after Objective Lineage.
Why it matters / Why this matters
It treats division of work through responsibility rather than conversation volume.
- Specialization could be represented as responsibility, authority, and state continuity rather than as a prompt persona.
- Wrong-target changes and hidden dependencies on another lane’s source could be detected as boundary failures.
The status marks the position of the canonical record; it is not a success claim.
Evidence / provenance
context:20260912-050541:task.laneIndexapplication:v6:application-origin-target-enforcement-rendezvous:20260829machine:v13:semantic-participants-separate-from-mutation-authority-v0:20260901mirage:v13:lane-local-validation-must-not-require-application-source:v0:20260904context:20260913-201241:public/development/index.jsongit:branch:lane/machine/20260816t021523z-e5d20dfd