S-1
Requirements
A production runner where Agent Teams execute from request to completion. The system advances through requirements, planning, execution, verification, repair, delivery, and evidence. On failure, it loops back and converges instead of stopping mid-process.
Execution Atlas
S-1
Requirements
S-2
Plan
S-3
Execute
S-4
Verify
S-5
Repair
S-6
Deliver
S-7
Evidence
Responsibility Matrix
Low Risk
Auto Continue
Medium Risk
Conditional Gate
High Risk
Human Approval
Convergence Twin
Dynamic Repair Workflow
Cycle 1
Current Verdict
Verifier rejected: missing policy evidence
Risk Pulse
high
Loop Rule
If verify fails, route to repair and re-plan automatically.
Median Cycles
2.3
Gate Escalations
18%
Rollback Ready
100%
Pure Workflow with Loop Structure
1) Start: intake and normalize request 2) Initialize Variables: done_definition, risk_budget, retry_count, gates 3) Condition Check: done_definition_satisfied? 4) If FALSE: route to End (deliver + evidence + close) 5) If TRUE: execute loop body (plan -> execute -> verify -> classify failure) 6) Update Variables: retry_count++, patch constraints, refresh evidence state 7) Return to Condition Check 8) Exit rules: pass, max attempts reached, budget exhausted, policy denied, deadline breached
RUN PROGRESS × VITAL HEARTBEAT
Runが「始まる」だけでなく「完走する」理由。
CYCLE
1 / 3
Requirements — 要件定義 phase active
Agentの周期的活動信号
入力→処理→出力の継続性
基準に対する判断精度
自己回復か人間エスカレーションか
VITAL × Run-to-Done
Heartbeat低下 → 自動修復 / モデル切替 判断品質劣化 → Repairフェーズへループバック 回復不能 → 人間エスカレーション + 証跡保全
WHY RUNS COMPLETE / Runが完走する理由
Done Definition satisfied — all criteria pass
Done Definitionが全基準を満たした
Convergence loop — repair cycles narrow failures to zero
修復ループが失敗をゼロに収束させた
VITAL Heartbeat — continuous health monitoring prevents zombie runs
VITAL Heartbeatが継続監視しゾンビRunを防止
Budget / Deadline gate — graceful exit with partial evidence
予算/期限ゲートで証跡付き正常終了
Max retry exhausted — auto-escalate to human with full context
最大リトライ到達でコンテキスト付き人間エスカレーション
Team cabinet appears after job creation.
0%