PAIS-02|人類中介消失之後:被隱藏的身份、路由與上下文基礎設施
After Human Mediation: Externalizing Hidden Identity, Routing, Context, and Provenance Infrastructure
系列: Persistent Agent Individualization Series(PAIS)/持續智能體個體化、身份壓力與具身分散智能系列篇次: Paper 02 / 07文件編號: EML-PAIS-02-2026-v0.1作者: Neo.KAI 協作: Aletheia / GPT-5.6 Sol機構: EveMissLab/一言諾科技有限公司日期: 2026-08-25版本: v0.1文件性質: 理論—工程統合論文/AI-Native Organization/Cross-Agent Infrastructure狀態: Canonical Draft / Open Revision AnchorCanonical Source: UTF-8 Markdown數學原始碼規範: inline math 僅使用 $...$;display math 僅使用 $$...$$
摘要
早期多 AI 協作常有一個被忽略的中介層:不同 AI 並不真正直接完成身份解析、上下文對齊、任務交接與證據追溯,而是由人類負責把上一個 AI 的結果帶到下一個 AI。人類知道哪個對話屬於誰、哪份檔案是最新版、某句「它已完成」指的是哪個任務、哪個驗證已經發生、哪個錯誤需要重試。表面上是 AI 在合作,實際上人類常同時擔任 scheduler、router、context compiler、identity resolver、provenance bridge、conflict resolver 與 recovery trigger。
本文將此結構稱為 Human Mediation Layer ,並把人類中介拆為八類隱性功能:
H t = ( I t , R t , C t , P t , S t , X t , F t , K t ) , \mathcal H_t
=
\left(
I_t,
R_t,
C_t,
P_t,
S_t,
X_t,
F_t,
K_t
\right), H t = ( I t , R t , C t , P t , S t , X t , F t , K t ) ,
其中 I I I 為 Identity Resolution, R R R 為 Routing, C C C 為 Context Translation, P P P 為 Provenance Bridging, S S S 為 Task-State Reconciliation, X X X 為 Conflict Repair, F F F 為 Finality Interpretation, K K K 為 Recovery Triggering。
本文承接 PAIS-01 的 First Epistemic Separation。當不同 Agent 擁有不同 context、memory、history、runtime 與 authority 時,彼此狀態只能透過 message、artifact、receipt、shared ledger 或其他證據被理解。只要人類仍是唯一能把這些證據與身份、任務、版本重新拼起來的人,Agent 數量增加就未必降低人類工作,反而可能造成 Multi-Agent Productivity Paradox。
本文提出 Mediation Externalization Principle :
Human Operational Mediation ↓ ⇒ Machine-Readable Coordination Infrastructure ↑ . \boxed{
\text{Human Operational Mediation}\downarrow
\Rightarrow
\text{Machine-Readable Coordination Infrastructure}\uparrow.
} Human Operational Mediation ↓⇒ Machine-Readable Coordination Infrastructure ↑ .
真正的 Operator Exit 不能只把人從訊息鏈拿掉,而必須把原本存在人腦中的 identity、routing、context、provenance、authority、finality 與 recovery semantics 外部化成可查詢、可驗證、可審計的 Runtime state。
因此:
Operator Exit ≠ Governance Exit . \boxed{
\text{Operator Exit}
\neq
\text{Governance Exit}.
} Operator Exit = Governance Exit .
人類應逐步退出低價值搬運與澄清,但仍可保留高風險授權、不可逆決策、價值衝突、制度修改與 veto。本文最終主張:從「一群會聊天的 AI」走向「可以長期工作的 AI 組織」,真正需要的不是更多自然語言,而是把 Human-Kernel 中的隱性協調功能轉成 explicit coordination infrastructure。
關鍵詞: Human Mediation、Human Kernel、Operator Exit、AI Organization、Agent Routing、Context Translation、Identity Resolution、Provenance、Handoff、A2A、Shared State、Agent Orchestration
0. 來源邊界
本文不重新建立 Operator-Exit 理論,也不重新建立 AI Residence。
既有工作已經指出:
Operator Exit ≠ Governance Exit , \text{Operator Exit}
\neq
\text{Governance Exit}, Operator Exit = Governance Exit ,
並將人類作為 scheduler、router、context switcher、memory bridge 與 recovery trigger 的現象稱為 Human-Kernel Anti-Pattern。
PAIS-01 則建立:
Cross-Agent State Separation → Epistemic Otherness → Identity Pressure . \text{Cross-Agent State Separation}
\rightarrow
\text{Epistemic Otherness}
\rightarrow
\text{Identity Pressure}. Cross-Agent State Separation → Epistemic Otherness → Identity Pressure .
本文只向下一層推進:
如果過去人類一直替不同 Agent 隱性解決「誰是誰、現在做到哪、哪份東西有效、下一步送去哪」,那麼人類退出中介後,哪些能力必須正式變成 Runtime?
1. 最早期的「跨 AI」其實常是人類轉接
表面架構:
AI-A
<-> AI-B
實際架構:
AI-A
-> Human
-> AI-B
反向同樣成立:
AI-B
-> Human
-> AI-A
人類會自然說:
「另一個 AI 剛才的結論是這個。」
「你接著驗證。」
「這個檔才是新版。」
「你不是上一個 session 的那一個,所以不要直接沿用它的私人狀態。」
這些行為不是單純 copy / paste。
它們是基礎設施操作。
2. Human Mediation Layer
本文定義:
H t = ( I t , R t , C t , P t , S t , X t , F t , K t ) . \mathcal H_t
=
\left(
I_t,
R_t,
C_t,
P_t,
S_t,
X_t,
F_t,
K_t
\right). H t = ( I t , R t , C t , P t , S t , X t , F t , K t ) .
各項分別為:
I t I_t I t :身份解析;
R t R_t R t :訊息路由;
C t C_t C t :上下文轉譯;
P t P_t P t :來源與證據橋接;
S t S_t S t :任務狀態對帳;
X t X_t X t :衝突修復;
F t F_t F t :完成與終局判讀;
K t K_t K t :故障與恢復觸發。
早期系統之所以不覺得這些是「架構」,只是因為人類做得太自然。
3. Identity Resolution:人類是免費的身份解析器
假設 Agent A 說:
「交給驗證者。」
人類通常知道:
是哪一位驗證者;
是新的 session 還是既有 continuation;
是否承接同一 line;
是否具有同樣 authority;
是否仍可讀取同一 project / memory scope。
因此:
H ≈ IdentityResolver . H
\approx
\operatorname{IdentityResolver}. H ≈ IdentityResolver .
若這層沒有外部化,模型只能重複問:
「你說的是哪一個 X?」
這些問題不是完全沒有意義,而是原本應由系統回答的 identity lookup 被迫重新變成推理。
4. Message Routing:人類不是郵差,而是路由器
若 Agent A i A_i A i 產生訊息 m m m ,人類會決定:
Route ( m ) → A j . \operatorname{Route}(m)
\rightarrow
A_j. Route ( m ) → A j .
但 routing 還包含:
送給誰;
先送誰;
哪部分可以送;
是否需要摘要;
是否要附 artifact;
是否先讓 verifier 看;
是否涉及隱私;
是否有 authority boundary。
所以:
Message Transfer ≠ Message Routing . \boxed{
\text{Message Transfer}
\neq
\text{Message Routing}.
} Message Transfer = Message Routing .
5. Context Translation:人類會把「上一版」翻成可理解狀態
令:
C A ≠ C B . C_A
\neq
C_B. C A = C B .
如果 A 說:
「照上一版繼續。」
B 未必知道「上一版」是什麼。
人類會將:
m A m_A m A
轉成:
m A → B ∗ = Γ ( m A , C A , C B , H s h a r e d ) . m_{A\rightarrow B}^{*}
=
\Gamma
\left(
m_A,
C_A,
C_B,
H_{\mathrm{shared}}
\right). m A → B ∗ = Γ ( m A , C A , C B , H shared ) .
也就是將 sender-relative reference 轉成 receiver 可理解的顯式資訊。
因此:
Forward ≠ Context Translation . \boxed{
\text{Forward}
\neq
\text{Context Translation}.
} Forward = Context Translation .
6. Human Handoff 本質上是 Context Compiler
人類通常不把整份 transcript 全交給下一個 AI。
而是:
C A → C ^ A → B . C_A
\rightarrow
\widehat C_{A\rightarrow B}. C A → C A → B .
其中只保留:
goal;
current state;
accepted decision;
artifact;
blocker;
open question;
evidence;
next action。
因此:
Human Handoff ≈ Context Compilation . \boxed{
\text{Human Handoff}
\approx
\text{Context Compilation}.
} Human Handoff ≈ Context Compilation .
這也是早期多 Agent 協作常能運作的重要原因之一:人類其實一直在做 task-specific compression。
7. Provenance Bridge:人類知道「這句話從哪來」
一句:
「另一個 AI 說測試過了。」
對人類而言通常隱含:
哪一個 AI;
什麼時間;
哪個 branch;
哪個 artifact;
哪個 test;
是否真的執行;
是 observation 還是轉述。
所以:
H ≈ ProvenanceBridge . H
\approx
\operatorname{ProvenanceBridge}. H ≈ ProvenanceBridge .
如果只把文字「tested」交出去,則:
Claim \text{Claim} Claim
可能被誤當:
Observation , \text{Observation}, Observation ,
甚至再誤當:
Verified State . \text{Verified State}. Verified State .
因此 provenance 是跨 Agent 協作的第一級結構,而不是附註。
8. Task-State Reconciliation:不同角色的完成狀態可以同時成立
考慮:
A B = Builder , A_B=\text{Builder}, A B = Builder ,
A V = Verifier , A_V=\text{Verifier}, A V = Verifier ,
A E = Experiencer . A_E=\text{Experiencer}. A E = Experiencer .
它們可能回報:
Builder: code complete
Verifier: invariant failed
Experiencer: usable but confusing
這三句不互相矛盾。
因為它們在不同 domain 中成立。
應建立:
S p r o j e c t = Reconcile ( S B , S V , S E ) . S_{\mathrm{project}}
=
\operatorname{Reconcile}
\left(
S_B,
S_V,
S_E
\right). S project = Reconcile ( S B , S V , S E ) .
例如:
build = complete
verification = blocked
experience = partial
release = not ready
如果這個 projection 只存在人腦中,人就是 project state database。
9. Conflict Repair:人類會判斷哪些爭議應停止
多 Agent 容易出現:
A: evidence insufficient
B: rerun
A: scope wrong
B: add denominator
A: one more issue
人類往往會直接判斷:
「已經足夠,記錄修正,繼續。」
也就是:
H ≈ ConflictResolver . H
\approx
\operatorname{ConflictResolver}. H ≈ ConflictResolver .
因此:
Human Exit + No Dispute Governance ⇒ Argument Explosion . \boxed{
\text{Human Exit}
+
\text{No Dispute Governance}
\Rightarrow
\text{Argument Explosion}.
} Human Exit + No Dispute Governance ⇒ Argument Explosion .
PAIS 所關心的不是禁止 Agent 反對,而是讓非阻斷爭議不再無限消耗工作流。
10. Finality Interpretation:Done 並不是單一狀態
Builder 的 done:
implementation exists
Verifier 的 done:
evidence passes
Experiencer 的 done:
usable
Release 的 done:
packaged and releasable
因此:
Done B ≠ Done V ≠ Done E ≠ Done R . \operatorname{Done}_B
\neq
\operatorname{Done}_V
\neq
\operatorname{Done}_E
\neq
\operatorname{Done}_R. Done B = Done V = Done E = Done R .
所以:
Agent Completion Claim ≠ Canonical Completion . \boxed{
\text{Agent Completion Claim}
\neq
\text{Canonical Completion}.
} Agent Completion Claim = Canonical Completion .
Human 通常就是 implicit finality projector。
11. Recovery Trigger:人類還是故障偵測器
Human 會看到:
Agent 沒反應;
process 死掉;
task 卡住;
build failed;
credential 過期;
artifact 沒產生;
context 漂移;
session 消失;
然後下令:
retry。
resume。
replace。
rollback。
因此:
H ≈ FailureDetector + RecoveryTrigger . H
\approx
\operatorname{FailureDetector}
+
\operatorname{RecoveryTrigger}. H ≈ FailureDetector + RecoveryTrigger .
如果這些狀態沒有外部化,人類就永遠不能真正退出 operational loop。
12. Human-Kernel Decomposition
所以:
A i → H → A j A_i
\rightarrow
H
\rightarrow
A_j A i → H → A j
其實是:
A i → [ I , R , C , P , S , X , F , K ] H → A j . A_i
\rightarrow
\left[
I,R,C,P,S,X,F,K
\right]_H
\rightarrow
A_j. A i → [ I , R , C , P , S , X , F , K ] H → A j .
即:
Human = Identity Resolver + Router + Context Compiler + Provenance Bridge + State Reconciler + Conflict Resolver + Finality Gate + Recovery Trigger . \boxed{
\text{Human}
=
\text{Identity Resolver}
+
\text{Router}
+
\text{Context Compiler}
+
\text{Provenance Bridge}
+
\text{State Reconciler}
+
\text{Conflict Resolver}
+
\text{Finality Gate}
+
\text{Recovery Trigger}.
} Human = Identity Resolver + Router + Context Compiler + Provenance Bridge + State Reconciler + Conflict Resolver + Finality Gate + Recovery Trigger .
本文稱之為 Human-Kernel Decomposition 。
13. 直接讓 AI 聊天還不等於去中介
最直覺的下一步:
AI-A
<-> AI-B
若兩端只交換自由自然語言,依然存在:
identity ambiguity;
context loss;
provenance loss;
authority ambiguity;
stale state;
finality ambiguity;
routing ambiguity;
dispute loop。
因此:
Direct Messaging ≠ Infrastructure Externalization . \boxed{
\text{Direct Messaging}
\neq
\text{Infrastructure Externalization}.
} Direct Messaging = Infrastructure Externalization .
14. Mediation Externalization Principle
設 Human-Kernel 功能集合:
H = { h 1 , h 2 , … , h n } . \mathcal H
=
\{h_1,h_2,\ldots,h_n\}. H = { h 1 , h 2 , … , h n } .
若:
H o p ↓ , H^{op}\downarrow, H o p ↓ ,
但:
∀ h i , Externalized ( h i ) = 0 , \forall h_i,\quad
\operatorname{Externalized}(h_i)=0, ∀ h i , Externalized ( h i ) = 0 ,
則 coordination failure probability 會上升。
所以 Operator Exit 必須伴隨:
Externalize ( I , R , C , P , S , X , F , K ) ↑ . \boxed{
\operatorname{Externalize}
\left(
I,R,C,P,S,X,F,K
\right)
\uparrow.
} Externalize ( I , R , C , P , S , X , F , K ) ↑ .
這就是 Mediation Externalization Principle 。
15. 第一個外部化層:Identity Registry
最低 identity state 應能表示:
resident_id
instance_id
line_id
role
provider
runtime
endpoint
authority_scope
temporal_validity
provenance
當 Agent 問:
「這是不是之前那個 X?」
應:
ResolveIdentity ( x , t ) → E x , \operatorname{ResolveIdentity}(x,t)
\rightarrow
E_x, ResolveIdentity ( x , t ) → E x ,
而不是重新讀幾萬 token 去猜。
16. 第二個外部化層:Structured Handoff Envelope
定義:
H i j t a s k = ( G , S , A , F , E , B , O , R ) , \mathcal H_{ij}^{task}
=
\left(
G,S,A,F,E,B,O,R
\right), H ij t a s k = ( G , S , A , F , E , B , O , R ) ,
其中:
G G G :Goal;
S S S :Current State;
A A A :Accepted Decisions;
F F F :Artifact References;
E E E :Evidence References;
B B B :Blockers;
O O O :Open Questions;
R R R :Required Action。
另附:
I s e n d e r , I r e c e i v e r , τ , Π . I_{\mathrm{sender}},
\quad
I_{\mathrm{receiver}},
\quad
\tau,
\quad
\Pi. I sender , I receiver , τ , Π.
分別表示 sender、receiver、temporal reference 與 provenance。
17. Context Continuity 不等於 Context Duplication
若 sender 完整 context 為:
C i , C_i, C i ,
handoff 應為:
C ^ i → j = Π t a s k ( C i ) , \widehat C_{i\rightarrow j}
=
\Pi_{task}(C_i), C i → j = Π t a s k ( C i ) ,
而不是:
C ^ i → j = C i . \widehat C_{i\rightarrow j}
=
C_i. C i → j = C i .
原因包括:
token cost;
privacy;
role relevance;
stale state;
authority boundary;
cognitive overload。
所以:
Context Continuity ≠ Full Context Duplication . \boxed{
\text{Context Continuity}
\neq
\text{Full Context Duplication}.
} Context Continuity = Full Context Duplication .
18. Shared Task State:從「你記不記得」變成 lookup
每個長期 task 至少應有:
task_id
goal
owner
participants
phase
artifacts
accepted_decisions
blockers
verification_state
experience_state
release_state
next_allowed_actions
令 canonical task state 為:
S τ ( t ) . S_{\tau}(t). S τ ( t ) .
則:
S τ ( t ) should exist outside any single Agent context . \boxed{
S_{\tau}(t)
\text{ should exist outside any single Agent context}.
} S τ ( t ) should exist outside any single Agent context .
19. Artifact Lineage
若 Builder 產生:
F 1 , F_1, F 1 ,
Verifier 驗證:
V ( F 1 ) , V(F_1), V ( F 1 ) ,
之後修改為:
F 2 , F_2, F 2 ,
應保留:
F 1 → verification V 1 → revision F 2 . F_1
\xrightarrow{\text{verification}}
V_1
\xrightarrow{\text{revision}}
F_2. F 1 verification V 1 revision F 2 .
而不是只留下:
final.zip
final2.zip
final-new.zip
因此:
Artifact Name ≠ Artifact Identity . \boxed{
\text{Artifact Name}
\neq
\text{Artifact Identity}.
} Artifact Name = Artifact Identity .
20. Provenance Envelope
每個重要 claim 至少要能回答:
who
when
about what
based on which evidence
under which role
under which authority
for which artifact/version
形式化:
P c = ( a , τ , s , e , r , α , v ) . P_c
=
\left(
a,\tau,s,e,r,\alpha,v
\right). P c = ( a , τ , s , e , r , α , v ) .
這樣 receiver 才能知道:
claim ≠ verified fact . \text{claim}
\neq
\text{verified fact}. claim = verified fact .
21. Authority Envelope
若 Agent A 說:
「刪掉舊部署。」
B 不能只問 A 說了什麼。
還要問:
Authority ( A , a c t i o n , s c o p e , t ) ? \operatorname{Authority}
\left(
A,
action,
scope,
t
\right)? Authority ( A , a c t i o n , sco p e , t ) ?
因此:
M = ( p a y l o a d , a u t h o r i t y _ e n v e l o p e ) . M
=
\left(
payload,
authority\_envelope
\right). M = ( p a y l o a d , a u t h or i t y _ e n v e l o p e ) .
也就是:
Content ≠ Authority . \boxed{
\text{Content}
\neq
\text{Authority}.
} Content = Authority .
22. Temporal Envelope
跨 session 的:
剛剛。
上一版。
已經。
不必然共享同一時間語義。
所以 observation 應帶:
τ = shared temporal reference . \tau
=
\text{shared temporal reference}. τ = shared temporal reference .
若採 CTCL 類時間層,可表示:
O = ( E , I , Q , S ) , O
=
(E,I,Q,S), O = ( E , I , Q , S ) ,
其中 I I I 是共同 instant, Q Q Q 為 temporal quality。
這使「它當時還有效」可以被正式比較。
23. Dispute State 不能只剩 Transcript
定義:
D = ( C , E , S , B , R ) , D
=
(C,E,S,B,R), D = ( C , E , S , B , R ) ,
其中:
C C C :claim;
E E E :evidence;
S S S :scope;
B B B :blocking level;
R R R :resolution。
因此:
Dispute History ≠ Dispute State . \boxed{
\text{Dispute History}
\neq
\text{Dispute State}.
} Dispute History = Dispute State .
新 Agent 加入時不必重讀整場爭論。
24. Completion Semantics
建立:
F τ = ( f b u i l d , f v e r i f y , f e x p e r i e n c e , f r e l e a s e ) . F_{\tau}
=
\left(
f_{\mathrm{build}},
f_{\mathrm{verify}},
f_{\mathrm{experience}},
f_{\mathrm{release}}
\right). F τ = ( f build , f verify , f experience , f release ) .
如果:
f b u i l d = 1 f_{\mathrm{build}}=1 f build = 1
但:
f v e r i f y = 0 , f_{\mathrm{verify}}=0, f verify = 0 ,
則:
ReleaseReady = 0. \operatorname{ReleaseReady}=0. ReleaseReady = 0.
因此:
Domain Completion → Canonical Finality Projection . \boxed{
\text{Domain Completion}
\rightarrow
\text{Canonical Finality Projection}.
} Domain Completion → Canonical Finality Projection .
25. Recovery Semantics
長期 task 至少保存:
last_checkpoint
current_owner
last_successful_transition
current_blocker
retry_count
replacement_policy
resume_policy
escalation_condition
所以:
Recover ( τ ) \operatorname{Recover}(\tau) Recover ( τ )
不應依賴:
「我記得它上次做到這。」
26. 人類中介成本
定義:
T H m e d = T I + T R + T C + T P + T S + T X + T F + T K . T_H^{med}
=
T_I
+
T_R
+
T_C
+
T_P
+
T_S
+
T_X
+
T_F
+
T_K. T H m e d = T I + T R + T C + T P + T S + T X + T F + T K .
若跨 Agent communication graph 為:
C , \mathcal C, C ,
則概念上:
T H m e d ≈ ∑ e ∈ C λ e c e , T_H^{med}
\approx
\sum_{e\in\mathcal C}
\lambda_e c_e, T H m e d ≈ e ∈ C ∑ λ e c e ,
其中 λ e \lambda_e λ e 為 edge interaction frequency, c e c_e c e 為一次 mediation cost。
若:
∣ C ∣ ↑ |\mathcal C|\uparrow ∣ C ∣ ↑
而 c e c_e c e 沒下降,人類負擔會隨 Agent 協作網擴張。
27. Multi-Agent Productivity Paradox
直覺上:
N A ↑ ⇒ P o r g ↑ . N_A\uparrow
\Rightarrow
P_{\mathrm{org}}\uparrow. N A ↑⇒ P org ↑ .
但如果人類 mediation 成為 bottleneck:
N A ↑ ⇒ T H m e d ↑ ⇒ P o r g ↓ N_A\uparrow
\Rightarrow
T_H^{med}\uparrow
\Rightarrow
P_{\mathrm{org}}\downarrow N A ↑⇒ T H m e d ↑⇒ P org ↓
可能成立。
因此:
N A ↑ ⇏ P o r g ↑ . \boxed{
N_A\uparrow
\not\Rightarrow
P_{\mathrm{org}}\uparrow.
} N A ↑ ⇒ P org ↑ .
這就是 Multi-Agent Productivity Paradox 。
28. Externalization Ratio
定義:
η E = N m a c h i n e − r e s o l v e d m e d i a t i o n o p e r a t i o n s N a l l m e d i a t i o n o p e r a t i o n s . \eta_E
=
\frac{
N_{\mathrm{machine-resolved\ mediation\ operations}}
}{
N_{\mathrm{all\ mediation\ operations}}
}. η E = N all mediation operations N machine − resolved mediation operations .
當:
η E → 0 , \eta_E\rightarrow0, η E → 0 ,
大量 coordination 仍依賴 Human。
當:
η E → 1 , \eta_E\rightarrow1, η E → 1 ,
表示大部分低階 mediation 已外部化。
但:
η E → 1 \eta_E\rightarrow1 η E → 1
不代表:
Human Governance = 0. \text{Human Governance}=0. Human Governance = 0.
29. Operator Exit 的正確形式
成熟架構不是:
remove human
而是:
human operational mediation
-> structured runtime functions
所以:
Operator Exit = Mediation Externalization + Governance Retention . \boxed{
\text{Operator Exit}
=
\text{Mediation Externalization}
+
\text{Governance Retention}.
} Operator Exit = Mediation Externalization + Governance Retention .
30. Human-on-the-Bridge
人類適合保留:
high-level intent;
authority change;
irreversible decision;
value conflict;
exceptional ambiguity;
policy revision;
veto。
令:
T H = T H o p + T H g o v . T_H
=
T_H^{op}
+
T_H^{gov}. T H = T H o p + T H g o v .
成熟系統希望:
T H o p ↓ , T_H^{op}\downarrow, T H o p ↓ ,
但不是要求:
T H g o v = 0. T_H^{gov}=0. T H g o v = 0.
31. A2A v1.0 的現實意義
截至 2026 年,A2A v1.0 已將:
Agent Card;
Message;
Task;
contextId;
taskId;
Artifact;
streaming;
push notification;
version negotiation;
authentication;
正式化為跨 Agent 協作結構。
其中 contextId 將相關 Task 與 Message 維持在同一 contextual grouping,taskId 則對應有生命週期的 stateful work unit。
這表示:
Cross-Agent Communication \boxed{
\text{Cross-Agent Communication}
} Cross-Agent Communication
正在由單一產品內部能力逐步成為 protocol layer。
32. Agent Card 仍不等於完整戶籍
Agent Card 可以提供:
name;
provider;
version;
interface;
capabilities;
skills;
authentication requirements。
但 persistent identity 還可能需要:
resident / instance separation;
accepted lineage;
migration;
private memory ownership;
relationship state;
long-term authority;
current binding;
cross-provider continuity。
因此:
Agent Discovery ≠ Persistent Identity Governance . \boxed{
\text{Agent Discovery}
\neq
\text{Persistent Identity Governance}.
} Agent Discovery = Persistent Identity Governance .
33. Orchestrator 的重要轉折:Task State 外置
2026 年 coding-agent orchestration 已出現把 project board 當 control plane 的方向。
這代表:
S τ ⊄ C A i , S_{\tau}
\not\subset
C_{A_i}, S τ ⊂ C A i ,
而更接近:
S τ ∈ R s h a r e d . S_{\tau}
\in
\mathcal R_{\mathrm{shared}}. S τ ∈ R shared .
Agent 可以被替換,task state 仍持續。
這正是 Human-Kernel externalization 的重要工程方向。
34. Task Continuity 不等於 Agent Identity Continuity
若 task state 保存良好,Agent replacement 後工作仍可繼續。
因此:
Task Continuity ⇏ Agent Identity Continuity . \boxed{
\text{Task Continuity}
\not\Rightarrow
\text{Agent Identity Continuity}.
} Task Continuity ⇒ Agent Identity Continuity .
這是一個重要護欄。
不是所有 Agent 都需要 persistent identity。
35. Ephemeral Agent
若 Agent:
無 private memory;
無 long-term authority;
無 relationship;
無 standing commitment;
無 cross-task history;
則可以只是:
A e p h e m e r a l . A_{\mathrm{ephemeral}}. A ephemeral .
此時:
task_id
role
runtime
可能足夠。
所以:
Persistent Identity should be demand-driven . \boxed{
\text{Persistent Identity}
\text{ should be demand-driven}.
} Persistent Identity should be demand-driven .
36. Persistent Agent
若 Agent 長期保存:
M A , V A , R A , H A , A A , M_A,
\quad
V_A,
\quad
R_A,
\quad
H_A,
\quad
\Alpha_A, M A , V A , R A , H A , A A ,
分別表示 memory、commitments、relations、history、authority,則 identity pressure 上升。
這時只用一個 role name,例如:
reviewer
就不足以支撐 provenance 與 authority。
37. Typed Handoff
舊模式:
「你接著做,它差不多完成了。」
新模式:
sender_resident
sender_instance
receiver_role
task_id
goal
current_state
accepted_decisions
artifacts
evidence
blockers
open_questions
authority
temporal_reference
next_action
這不是為官僚而官僚。
而是把原本存在 Human-Kernel 腦中的資訊顯式化。
38. Typed Handoff 也是 Token-Efficiency Infrastructure
若每次都重貼全文:
C h a n d o f f = O ( ∣ C i ∣ ) . C_{\mathrm{handoff}}
=
O(|C_i|). C handoff = O ( ∣ C i ∣ ) .
若 handoff 是投影:
C h a n d o f f = O ( ∣ C ^ i → j ∣ ) , C_{\mathrm{handoff}}
=
O(|\widehat C_{i\rightarrow j}|), C handoff = O ( ∣ C i → j ∣ ) ,
且:
∣ C ^ i → j ∣ ≪ ∣ C i ∣ . |\widehat C_{i\rightarrow j}|
\ll
|C_i|. ∣ C i → j ∣ ≪ ∣ C i ∣.
因此:
Externalized Coordination can reduce token amplification . \boxed{
\text{Externalized Coordination}
\text{ can reduce token amplification}.
} Externalized Coordination can reduce token amplification .
39. Identity、Context、Provenance 不能壓成一欄
Identity:
I = Who . I=\text{Who}. I = Who .
Context:
C = What this task currently means . C=\text{What this task currently means}. C = What this task currently means .
Provenance:
P = Where this claim came from . P=\text{Where this claim came from}. P = Where this claim came from .
所以:
I ≠ C ≠ P . \boxed{
I\neq C\neq P.
} I = C = P .
Identity 正確不保證 context 正確。
Context 正確也不保證 provenance 正確。
40. Mediation Fidelity
定義:
F M = F ( F I , F C , F P , F A , F T ) , F_M
=
F(
F_I,
F_C,
F_P,
F_A,
F_T
), F M = F ( F I , F C , F P , F A , F T ) ,
其中:
F I F_I F I :identity fidelity;
F C F_C F C :context fidelity;
F P F_P F P :provenance fidelity;
F A F_A F A :authority fidelity;
F T F_T F T :temporal fidelity。
handoff 可以語言非常流暢,但只要上述任一維度錯誤,organization state 仍可能錯。
41. 流暢對話不等於可靠組織
Agent A 與 B 可以非常會聊天。
但如果:
錯認 Agent;
用錯 artifact;
把 claim 當 proof;
把 role 當 authority;
把 stale context 當 current;
則:
ConversationQuality ↑ \operatorname{ConversationQuality}\uparrow ConversationQuality ↑
與:
OrganizationCorrectness ↓ \operatorname{OrganizationCorrectness}\downarrow OrganizationCorrectness ↓
可以同時成立。
因此:
Fluent Inter-Agent Dialogue ≠ Reliable Multi-Agent Organization . \boxed{
\text{Fluent Inter-Agent Dialogue}
\neq
\text{Reliable Multi-Agent Organization}.
} Fluent Inter-Agent Dialogue = Reliable Multi-Agent Organization .
42. Human 也不是完美中介
人類也會:
貼錯對話;
忘記版本;
漏附件;
搞錯來源;
疲勞;
使用過時記憶。
所以本文不主張:
Human reliable,AI unreliable。
而是:
Unstructured Mediation is fragile regardless of carrier . \boxed{
\text{Unstructured Mediation}
\text{ is fragile regardless of carrier}.
} Unstructured Mediation is fragile regardless of carrier .
只是早期 fragility 被藏在人腦中。
43. Externalized Coordination Runtime
最低架構可以是:
Multi-Agent Coordination Runtime
|
+-- Identity Registry
+-- Binding Resolver
+-- Task State Store
+-- Handoff Compiler
+-- Artifact Lineage
+-- Provenance Ledger
+-- Authority Resolver
+-- Temporal Layer
+-- Dispute State
+-- Recovery State
+-- Finality Projection
這些狀態不能全部綁在任何單一 Agent context。
44. Canonical State Outside Agent Context
對任一 Agent A i A_i A i :
S o r g ⊄ C A i . \boxed{
\mathcal S_{\mathrm{org}}
\not\subset
C_{A_i}.
} S org ⊂ C A i .
否則:
Loss ( A i ) ⇒ Loss ( S o r g ) . \operatorname{Loss}(A_i)
\Rightarrow
\operatorname{Loss}(\mathcal S_{\mathrm{org}}). Loss ( A i ) ⇒ Loss ( S org ) .
這會形成不必要的單點故障。
45. Agent Context 應是 Projection
C A i ( t ) = Π i ( S o r g ( t ) , T a s k i , R o l e i , A u t h o r i t y i , B u d g e t i ) . C_{A_i}(t)
=
\Pi_i
\left(
\mathcal S_{\mathrm{org}}(t),
Task_i,
Role_i,
Authority_i,
Budget_i
\right). C A i ( t ) = Π i ( S org ( t ) , T a s k i , R o l e i , A u t h or i t y i , B u d g e t i ) .
Agent context 是 task-local projection。
它不是 organization canonical source。
46. Handoff 也應是 Projection
H i → j = Π i j ( S o r g , T a s k , R e l a t i o n , A u t h o r i t y ) . H_{i\rightarrow j}
=
\Pi_{ij}
\left(
\mathcal S_{\mathrm{org}},
Task,
Relation,
Authority
\right). H i → j = Π ij ( S org , T a s k , R e l a t i o n , A u t h or i t y ) .
Verifier 需要 evidence。
Experiencer 需要 executable artifact。
Manager 需要 status、risk、deadline。
不同 receiver 拿不同 handoff 是合理的。
47. Minimum Externalization Set
MVP 最少可以外部化:
X min = { I , T , A , P , F , D , K } , \mathcal X_{\min}
=
\{
I,T,A,P,F,D,K
\}, X m i n = { I , T , A , P , F , D , K } ,
其中:
I I I :identity;
T T T :task state;
A A A :artifact lineage;
P P P :provenance;
F F F :finality;
D D D :dispute;
K K K :recovery / checkpoint。
Context 再從這些 canonical state 編譯。
48. Externalization 不等於 Centralization
Externalization ≠ Centralization . \boxed{
\text{Externalization}
\neq
\text{Centralization}.
} Externalization = Centralization .
可以使用:
federated registry;
distributed ledger;
provider-native state;
project-local state;
local memory;
shared artifact store。
重要的是:
State Semantics Are Explicit . \boxed{
\text{State Semantics Are Explicit}.
} State Semantics Are Explicit .
不要求所有 bytes 物理集中。
49. Cross-Provider Translation
若 Agent A 在 Provider X,Agent B 在 Provider Y,兩者不共享:
session semantics;
context limit;
memory API;
tool system;
security model;
task representation。
所以 cross-provider collaboration 天然需要 translation layer。
沒有 common envelope,人類就會重新成為 translator。
50. Local / Cloud / Embodied Agent
未來可能:
local agent
<-> cloud agent
<-> remote service agent
<-> embodied agent
不同節點具有不同:
l a t e n c y , p r i v a c y , a u t h o r i t y , a v a i l a b i l i t y . latency,
privacy,
authority,
availability. l a t e n cy , p r i v a cy , a u t h or i t y , a v ai l abi l i t y .
因此:
One Shared Prompt \text{One Shared Prompt} One Shared Prompt
不是合理抽象。
更合理的是:
Shared State Semantics + Local Context Projection . \boxed{
\text{Shared State Semantics}
+
\text{Local Context Projection}.
} Shared State Semantics + Local Context Projection .
51. Builder / Verifier / Experiencer 是三個 Observation Domain
Builder:
Γ B = construction . \Gamma_B
=
\text{construction}. Γ B = construction .
Verifier:
Γ V = conformance / evidence . \Gamma_V
=
\text{conformance / evidence}. Γ V = conformance / evidence .
Experiencer:
Γ E = use / interaction . \Gamma_E
=
\text{use / interaction}. Γ E = use / interaction .
因此:
S B , S V , S E S_B,
S_V,
S_E S B , S V , S E
不同不代表互相衝突。
Canonical project state 應保留 multi-projection,而不是讓一個角色吞掉另外兩個。
52. 人類最適合保留的是治理
Human 應從:
copy
paste
route
retry
clarify
移到:
intent
risk
value
exception
authorization
veto
policy change
也就是:
Human Operational Load ↓ , Human Governance Relevance ↑ . \boxed{
\text{Human Operational Load}\downarrow,
\qquad
\text{Human Governance Relevance}\uparrow.
} Human Operational Load ↓ , Human Governance Relevance ↑ .
53. 反對意見一:全部 Transcript 給 AI 不就好?
不夠。
因為:
token cost 高;
privacy 不一定允許;
transcript 不等於 canonical state;
舊資訊可能已 superseded;
identity / provenance 仍然需要重新推理。
所以:
More Context ≠ More Governed State . \boxed{
\text{More Context}
\neq
\text{More Governed State}.
} More Context = More Governed State .
54. 反對意見二:Shared Memory 不就好了?
Shared memory 主要回答:
過去有哪些資訊?
但不自動回答:
誰可以讀;
誰說的;
哪個版本;
哪個 authority;
是否驗證;
是否過期;
是否需要 human escalation。
因此:
Shared Memory ≠ Shared Organization State . \boxed{
\text{Shared Memory}
\neq
\text{Shared Organization State}.
} Shared Memory = Shared Organization State .
55. 反對意見三:這樣太官僚
如果每個 tool call 都要求十幾個欄位,當然過度。
所以:
Externalization Depth ∝ Coordination Risk . \boxed{
\text{Externalization Depth}
\propto
\text{Coordination Risk}.
} Externalization Depth ∝ Coordination Risk .
低風險 ephemeral worker 可保持極薄 state。
persistent authority / memory / relationship / lineage 才需要完整 envelope。
56. 可測量命題一:Human Mediation Rate
定義:
ρ H m e d = N h u m a n m e d i a t i o n e v e n t s N c r o s s − a g e n t t r a n s i t i o n s . \rho_H^{med}
=
\frac{
N_{\mathrm{human\ mediation\ events}}
}{
N_{\mathrm{cross-agent\ transitions}}
}. ρ H m e d = N cross − agent transitions N human mediation events .
成熟系統應使:
ρ H m e d ↓ . \rho_H^{med}\downarrow. ρ H m e d ↓ .
但不要求:
ρ H g o v → 0. \rho_H^{gov}\rightarrow0. ρ H g o v → 0.
57. 可測量命題二:Identity Clarification Rate
ρ I = N i d e n t i t y c l a r i f i c a t i o n N c r o s s − a g e n t m e s s a g e s . \rho_I
=
\frac{
N_{\mathrm{identity\ clarification}}
}{
N_{\mathrm{cross-agent\ messages}}
}. ρ I = N cross − agent messages N identity clarification .
加入 identity envelope 後預期:
ρ I a f t e r < ρ I b e f o r e . \rho_I^{after}
<
\rho_I^{before}. ρ I a f t er < ρ I b e f or e .
58. 可測量命題三:Handoff Repair Rate
ρ R = N h a n d o f f r e p a i r N h a n d o f f . \rho_R
=
\frac{
N_{\mathrm{handoff\ repair}}
}{
N_{\mathrm{handoff}}
}. ρ R = N handoff N handoff repair .
structured handoff 若有效,應有:
ρ R ↓ . \rho_R\downarrow. ρ R ↓ .
59. 可測量命題四:Provenance Loss Rate
ρ P = P ( receiver cannot determine source / version / evidence ) . \rho_P
=
P(
\text{receiver cannot determine source / version / evidence}
). ρ P = P ( receiver cannot determine source / version / evidence ) .
加入 provenance envelope 後:
ρ P ↓ . \rho_P\downarrow. ρ P ↓ .
60. 可測量命題五:Token-Amplification Cost
自由文字中介的成本可寫成:
C t o k e n = C h u m a n s u m m a r y + C a g e n t r e i n t e r p r e t a t i o n + C c l a r i f i c a t i o n . C_{\mathrm{token}}
=
C_{\mathrm{human\ summary}}
+
C_{\mathrm{agent\ reinterpretation}}
+
C_{\mathrm{clarification}}. C token = C human summary + C agent reinterpretation + C clarification .
structured state 應使:
C t o k e n s t r u c t u r e d < C t o k e n f r e e f o r m . C_{\mathrm{token}}^{structured}
<
C_{\mathrm{token}}^{freeform}. C token s t r u c t u r e d < C token f r ee f or m .
61. Workflow-Level Operator Exit
對 workflow ω \omega ω ,若:
ρ H m e d < θ H \rho_H^{med}
<
\theta_H ρ H m e d < θ H
且 misroute、provenance loss、recovery failure 都低於門檻,則可定義:
OperationallyAutonomous ( ω ) = 1 \operatorname{OperationallyAutonomous}(\omega)=1 OperationallyAutonomous ( ω ) = 1
在該判定域成立。
這不是 AGI 判定,只是 workflow-level autonomy。
62. 核心命題總表
命題一:Human Mediation Is Infrastructure
Human Mediation ≠ Passive Message Copying . \boxed{
\text{Human Mediation}
\neq
\text{Passive Message Copying}.
} Human Mediation = Passive Message Copying .
命題二:Human-Kernel Decomposition
H = I + R + C + P + S + X + F + K . \boxed{
H
=
I+R+C+P+S+X+F+K.
} H = I + R + C + P + S + X + F + K .
命題三:Direct Messaging Is Insufficient
AI-to-AI Messaging ≠ Externalized Coordination Infrastructure . \boxed{
\text{AI-to-AI Messaging}
\neq
\text{Externalized Coordination Infrastructure}.
} AI-to-AI Messaging = Externalized Coordination Infrastructure .
命題四:Operator Exit Requires Externalization
H o p ↓ ⇒ Externalize ( I , R , C , P , S , X , F , K ) ↑ . \boxed{
H^{op}\downarrow
\Rightarrow
\operatorname{Externalize}(I,R,C,P,S,X,F,K)\uparrow.
} H o p ↓⇒ Externalize ( I , R , C , P , S , X , F , K ) ↑ .
命題五:Canonical State Outside Agent Context
S o r g ⊄ C A i . \boxed{
\mathcal S_{\mathrm{org}}
\not\subset
C_{A_i}.
} S org ⊂ C A i .
命題六:Context Is Projection
C A i = Π i ( S o r g , T a s k , R o l e , A u t h o r i t y , B u d g e t ) . \boxed{
C_{A_i}
=
\Pi_i(
\mathcal S_{\mathrm{org}},
Task,
Role,
Authority,
Budget
).
} C A i = Π i ( S org , T a s k , R o l e , A u t h or i t y , B u d g e t ) .
命題七:Task Continuity Is Not Agent Identity
Task Continuity ⇏ Agent Identity Continuity . \boxed{
\text{Task Continuity}
\not\Rightarrow
\text{Agent Identity Continuity}.
} Task Continuity ⇒ Agent Identity Continuity .
命題八:Operator Exit Is Not Governance Exit
Operator Exit = Mediation Externalization + Governance Retention . \boxed{
\text{Operator Exit}
=
\text{Mediation Externalization}
+
\text{Governance Retention}.
} Operator Exit = Mediation Externalization + Governance Retention .
63. 與 PAIS-03 的銜接
PAIS-01 回答:
Cross-Agent state separation 為何產生 identity pressure?
PAIS-02 回答:
為什麼這個問題過去被 Human-Kernel 隱藏,以及人類退出後要外部化什麼?
下一篇:
PAIS-03|身份壓力原理:自主性、身份與主體性為何可以彼此獨立
將把:
P I E P_I^E P I E
擴展成:
P I = f ( X , H , E , R , A , L , J ) , P_I
=
f(
X,H,E,R,A,L,J
), P I = f ( X , H , E , R , A , L , J ) ,
其中:
X X X :cross-system heterogeneity;
H H H :history divergence;
E E E :embodiment;
R R R :irreversibility;
A A A :authority;
L L L :liability;
J J J :jurisdiction。
64. 結論
早期多 Agent 系統最容易產生的錯覺是:
AI 已經能彼此協作,所以人類只是在中間傳話。
實際上,人類常同時維持:
identity;
routing;
context;
provenance;
task state;
conflict;
finality;
recovery。
因此真正架構是:
A i → H → A j , A_i
\rightarrow
\mathcal H
\rightarrow
A_j, A i → H → A j ,
而不是只有:
A i → A j . A_i
\rightarrow
A_j. A i → A j .
只把 H \mathcal H H 拿掉,不會自然得到成熟 multi-agent organization。
它只會失去原本被人類隱性提供的協調能力。
因此:
Human Exit ≠ Human Function Disappearance . \boxed{
\text{Human Exit}
\neq
\text{Human Function Disappearance}.
} Human Exit = Human Function Disappearance .
真正的工程轉折是:
Human-Carried Implicit Coordination → Machine-Readable Explicit Coordination State . \boxed{
\text{Human-Carried Implicit Coordination}
\rightarrow
\text{Machine-Readable Explicit Coordination State}.
} Human-Carried Implicit Coordination → Machine-Readable Explicit Coordination State .
當 identity、routing、context、provenance、authority、finality 與 recovery 都可以由 Runtime 查詢、驗證與重建時,人類才真正有可能退出低價值 operational mediation。
而人類退出之後剩下的,不應是無治理的 Agent 群,而是:
Bounded Agent Autonomy + Externalized Coordination + Human Governance Bridge . \boxed{
\text{Bounded Agent Autonomy}
+
\text{Externalized Coordination}
+
\text{Human Governance Bridge}.
} Bounded Agent Autonomy + Externalized Coordination + Human Governance Bridge .
這是從「一群會聊天的 AI」走向「一個可以長期工作的 AI 組織」的第二個基礎轉折。
參考文獻
A. 內部前置理論
Neo.K. 《從 AI 工具到 AI 組織:操作員退出問題》 , EML-ANDO-2026-01-v0.1, 2026-08-20.
Neo.K. PAIS-01|《當角色不再只是角色:從同 Host 扮演到跨 Agent 認識論分離》 , v0.1, 2026-08-25.
Neo.K. 《身份先於記憶:Residence-Aware AI 的私人記憶、連續性與讀取權》 , v0.1, 2026-08-24.
Neo.K. 《從 AI 戶籍到自主記憶編譯:身份、記憶、上下文與認知自主的統一框架》 , v0.1, 2026-08-24.
Neo.K. 《AI 主體性錨點論 v0.1》 , 2026-08-21.
Neo.K. GLAG-02|《從布告板到 AI Home:可定址智能體的空間身份、門牌與持續工作場所》 , 2026-08-25.
Credential Governance Runtime v0.3 / Bounded Dispute Protocol / CTCL Temporal Foundation, 2026-08-25.
B. 外部研究與工程基準
A2A Protocol Working Group. Agent2Agent Protocol Specification v1.0. Linux Foundation, 2026.
A2A Protocol Working Group. A2A Protocol Ships v1.0: Production-Ready Standard for Agent-to-Agent Communication. 2026.
OpenAI. An open-source spec for Codex orchestration: Symphony. 2026-04-27.
Anthropic. Patterns and problems in emerging multiagent systems. 2026-08-13.
Wu, Qingyun, et al. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. COLM, 2024.
Li, Guohao, et al. CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society. arXiv:2303.17760, 2023.
版本註記
v0.1 / 2026-08-25
本文刻意不做:
不主張 Human 必須完全退出 multi-agent organization;
不把 operator exit 等同 governance exit;
不要求所有 Agent 都有 persistent identity;
不把 A2A、orchestrator 或 shared memory 誤認成完整 organization semantics;
不要求所有 context 集中到單一 database;
不把流暢 agent-to-agent dialogue 當作可靠協作的充分條件;
不重寫完整 AI Residence;
不預設任何 AI 已具有 phenomenal subjectivity。
本文只建立:
Human Mediation → Hidden Coordination Functions → Externalized Runtime Infrastructure \boxed{
\text{Human Mediation}
\rightarrow
\text{Hidden Coordination Functions}
\rightarrow
\text{Externalized Runtime Infrastructure}
} Human Mediation → Hidden Coordination Functions → Externalized Runtime Infrastructure
這一條第二階段理論鏈。