Runtime pattern
DAG + policy gates
Agent Execution Graph - bp Sphere
How specialized agents coordinate, hand off, recover, and produce replayable decisions.
Agent runtimeReplayableTechnology proof
Agent Execution Graph
This page answers the technology-workshop question: how do agents coordinate without becoming a black box? Every transition below represents a bounded runtime contract: input object, output object, policy gate, timing, failure path, and replay anchor.
Loop prevention
Correlation + TTL
Conflict handling
Policy arbitration
Replay support
Signal-to-action trace
Execution path
A mission starts from a signal and advances through a graph of specialist agents. Agents do not call each other freely; the runtime advances state only when the previous contract is complete.
| 1. Signal Agent | Classifies event, assigns severity, correlation id, mission. |
|---|---|
| 2. Context Agent | Resolves supplier, invoice, journal, entity, cost center, or exposure object. |
| 3. Evidence Agent | Builds source-backed evidence pack with transaction and document lineage. |
| 4. Policy Agent | Evaluates authority, SOX, approval, write-back, and data access controls. |
| 5. Impact Agent | Quantifies financial, operational, risk, and readiness impact. |
| 6. Decision Agent | Produces recommendation, confidence, alternatives, and human owner. |
| 7. Execution Agent | Drafts workflow, ticket, hold, communication, or write-back payload. |
| 8. Replay/Learning Agent | Seals trace, captures outcome, proposes governed learning. |
Runtime controls
The graph is designed for enterprise operation: bounded state changes, visible ownership, retry paths, and replayable outputs.
| Agent identity | Every execution uses a registered agent id, owner, version, permissions, and skills. |
|---|---|
| State contract | Each node declares required input, produced output, and completion criteria. |
| Timing view | Per-node latency, queue wait, retry count, and SLA pressure are tracked. |
| Dependency view | Required SOR adapters, policies, evidence sources, and skills are listed. |
| Loop guard | Correlation id, event TTL, duplicate suppression, and max-hop limits prevent loops. |
| Conflict guard | Policy runtime arbitrates incompatible recommendations or action conflicts. |
| Failure guard | Retry, fallback, dead-letter quarantine, and human escalation are explicit. |
| Replay anchor | Each node emits replay metadata: input hash, output hash, policy result, timestamp. |
Visual execution graph
Drill-down proof points
| Question | Where to prove it live |
|---|---|
| Which agents participated? | Agent Inventory + this execution graph |
| What happened if an agent failed? | Runtime Command Center failure and retry panels |
| Can the decision be replayed? | Replay Center and Replay + Learning Evolution |
| How are conflicts handled? | Policy Runtime arbitration and human governance route |