Skip to main content

ADR-004: Promise.allSettled for Phase 1 Agent Orchestration

Status: Accepted Date: 2026-02-25 Deciders: OpsFlow LLC (DTCC/FINOS Hackathon team)

Context​

CALMGuard runs 4 AI agents in two phases. Phase 1 agents (Architecture Analyzer, Compliance Mapper, Pipeline Generator, Cloud Infra Generator) are independent and can run in parallel. The orchestrator needed a concurrency strategy for Phase 1.

Options considered:

  • A) Sequential execution — safest, slowest; 4x LLM latency
  • B) Promise.all — parallel, but one agent failure cancels the entire analysis
  • C) Promise.allSettled — parallel with graceful degradation; each agent's success/failure is independent

Decision​

Use Promise.allSettled for Phase 1 parallel execution in src/lib/agents/orchestrator.ts. Each agent result is inspected individually after settlement. Failed agents contribute null to the AnalysisResult and are tracked in failedAgents[].

Consequences​

Good:

  • Wall-clock time reduced from ~4x to ~1x LLM latency for Phase 1 (agents run concurrently)
  • One agent's LLM timeout does not cancel the other 3 agents
  • Dashboard shows partial results when some agents succeed — users get value even on partial failure
  • failedAgents array in AnalysisResult provides transparency into what succeeded

Neutral:

  • Risk Scorer (Phase 2) requires Architecture + Compliance results. If both fail, Risk Scorer is skipped. This is the correct behavior.

Bad:

  • Higher peak LLM API cost (4 concurrent requests vs sequential). Acceptable given demo context.