AIチーム／AI組織論

複数agentを置くだけでは重複、責任境界消失、handoff context欠落、並列副作用、停止条件不在、cost/token増幅が起こる。将来版ではAIチームを擬人的役職表ではなくtask decomposition、ownership、message/artifact contract、authority、shared state、human escalation、evaluationを持つ運用構造として扱う。

Enterprise AI Core / agent architecture / organization overlay / interoperability / governance / implementation / evaluation。

multi-agentをsingle-agent/workflowより優先する条件は何か。role/authority/state/handoff/artifact/stop-escalationをどう記録するか。

• Anthropic, "How we built our multi-agent research system", 2025-06-13, accessed 2026-09-25, https://www.anthropic.com/engineering/multi-agent-research-system --- SOURCE_REPORT+PRODUCT_VOLATILE, support+limit。orchestrator-worker、internal research eval +90.2%、multi-agent約15x chat tokens、共有context/依存taskの不適合。 - Microsoft, "Workflow orchestrations in Agent Framework", accessed 2026-09-25, https://learn.microsoft.com/en-us/agent-framework/workflows/orchestrations/ --- VERIFIED_SPEC+PRODUCT_VOLATILE。Sequential/Concurrent/Handoff/Group Chat/Magentic、HITL。 - Microsoft Agent Framework, "Agent Framework's Orchestration Patterns Reach 1.0", 2026-07-08, accessed 2026-09-25, https://devblogs.microsoft.com/agent-framework/agent-frameworks-orchestration-patterns-reach-1-0/ --- VERIFIED_SPEC+PRODUCT_VOLATILE。Python agent-framework-orchestrations 1.0.0。 - A2A Protocol, "A2A Protocol Ships v1.0", 2026-03-12, accessed 2026-09-25, https://a2a-protocol.org/dev/blog/2026/03/12/a2a-protocol-ships-v10-production-ready-standard-for-agent-to-agent-communication/ --- VERIFIED_SPEC。v1.0、multi-protocol/version negotiation/multi-tenancy/signed Agent Cards、MCPとA2Aの層分離。 - A2A Protocol, "Core Concepts", accessed 2026-09-25, https://a2a-protocol.org/latest/topics/key-concepts/ --- VERIFIED_SPEC。Agent Card/Task/Message/Part/Artifact/contextId/auth。 - A2A Protocol, "Roadmap", updated 2026-09-15, accessed 2026-09-25, https://a2a-protocol.org/latest/roadmap/ --- VERIFIED_SPEC+PRODUCT_VOLATILE, limit。v1.1 task timeline/event filtering/working-state messages、BiDi streaming、elicitation/multi-turnはnear-term initiative。

AIチームは ORCHESTRATION_TOPOLOGY + ROLE_AUTHORITY + TASK_OWNERSHIP + STATE_OWNERSHIP + MESSAGE/ARTIFACT CONTRACT + HANDOFF + HUMAN_GATE + STOP/RECOVERY として定義する。patternごとの依存関係・並列性・control transfer・共有context量を分離する。Anthropicの+90.2%は特定internal research evalであり一般保証ではない。A2A適合だけでも業務責任・権限妥当性・品質は保証されない。人間/組織ownerとagent execution roleを分離する。

single-agent/決定論workflowで十分なtaskではcoordination、latency、token/cost、observability、recoveryが悪化し得る。race/duplicate side effect、handoff loss、circular delegation、manager bottleneck、shared-state contamination、authority escalation、partial completion、version driftをfailure classとする。

Microsoft patternはSDK/version/runtime依存。A2Aはv1.0 current stable baselineだがendpoint/auth/supportedInterfaces/extensions/tenant/network policyは実装別確認。Anthropic内部architectureは任意tenant/APIで同一構成を設定できることを意味しない。

PROPOSAL AI_TEAM_CONTRACT {team_id, objective, topology, agent_id, role, allowed_tasks, authority_scope, tools, data_scope, task_owner, state_owner, input_contract, output_artifact_contract, handoff_edges, approval_gates, concurrency_policy, idempotency_policy, stop_conditions, escalation_owner, retry_budget, cost_budget, trace_ids, version_manifest}。deterministic workflow→single agent+tools→bounded specialists/handoff→concurrent/managerの順に必要性が観測された場合だけ複雑化。

同一golden taskをbaselineと比較しaccepted completion、coverage、duplicate-work、handoff loss、wrong-owner routing、conflicting action、tool calls、tokens/cost、p50/p95 latency、human intervention、recovery successを測る。

least privilegeをagent単位に割当てshared credentialを既定にしない。外部writeはowner/approval/idempotency/reconciliationを要求。message/artifactへdata minimization/provenanceを適用しtask/handoff/tool actionをcorrelation可能にする。fan-out/token/cost上限、circular delegation検知、kill/escalation経路を持つ。

「複数AIを組み合わせれば自動的に性能が上がるわけではない。独立して並列化できる高価値taskではmulti-agentが有効な場合がある一方、共有contextや依存関係が多いtaskではcoordination costとfailure surfaceが増える。AIチームは役割・権限・state所有・handoff・成果物・停止条件を明示したcontractとして設計し、single-agentまたはworkflow baselineと比較して採用する。」

EA-006、EA-020、EA-035、EA-036とcross-reference。AI_TEAM_CONTRACTをEA-029 template packへ接続。publication時にA2A current versionを再pinし、provider mappingではplan/region/admin/tenant gateを別採集。

/ STOP_CONDITION: provider横断・同条件multi-agent vs single-agent benchmark、組織横断A2A incident/case study、shared-state ownership標準schema、法的accountability上のAI role表現はUNKNOWN。必要な証拠・反証・failure envelope・availability・評価・safe wordingが揃ったためEA-037はBOUNDED_CLOSE。

Microsoft Agent Framework orchestration patterns are 1.0 across Python/.NET at audit; A2A v1.0 stable baseline; Sep 15 roadmap v1.1 items remain near-term work. Anthropic +90.2% internal research-eval gain and ~15x chat-token cost remain provider-specific SOURCE_REPORT, not a general guarantee.

A2A conformance != business responsibility/authorization/quality; Microsoft/Anthropic patterns are not universal taxonomy; human/organization owner distinct from agent execution role.

preserve provider-cross benchmark, cross-organization incident/case study, shared-state ownership standard schema, legal accountability role expression.