Date: August 17, 2025 Status: Superseded (status corrected 2026-08-29) Deciders: Principal Architect, Chief Architect, Chief of Staff Classification: Architectural (Strategic)
Initial architectural decisions assumed orchestration would be reserved for complex, ambiguous tasks. The implementation of PM-033d’s enhanced autonomy (achieving 0ms coordination latency and 4+ hour autonomous operations) revealed that orchestration overhead has become negligible.
Combined with Chain-of-Draft’s 92% token reduction (ADR-016) and MCP’s federation capabilities (ADR-018), the economics of orchestration have fundamentally changed. What was once expensive coordination is now essentially free.
The phrase “going whole hog” emerged during MCP week when we realized: why maintain two code paths (direct and orchestrated) when orchestration adds no measurable overhead?
We commit to orchestration-first architecture for ALL Piper Morgan operations, not just complex or ambiguous tasks. Every operation, from simple file lookups to complex multi-agent workflows, will route through the orchestration layer.
“If it can be orchestrated, it should be orchestrated.”
This isn’t premature optimization—it’s architectural simplification. One code path, one pattern, infinite composability.
class OrchestrationFirst:
"""Every operation goes through orchestration."""
async def execute_any_task(self, task: Task) -> Result:
# No more if/else for simple vs complex
# Everything gets orchestrated
return await self.orchestrator.execute(task)
# Deleted code:
# if task.is_simple():
# return await direct_execution(task) # ❌ NO MORE
# else:
# return await orchestrator.execute(task)
User Request
↓
Orchestration Layer (ALWAYS)
├── Ambiguity Assessment
├── Token Optimization (CoD)
├── Agent Selection
├── Excellence Flywheel Verification
└── Kind Communication Wrapper
↓
Execution (Single or Multi-Agent)
↓
Systematic Validation
↓
Response
Approach: Orchestrate only when complexity score > 0.5 Why Rejected: Maintains dual code paths. Threshold tuning becomes endless debate. PM-033d proved overhead is negligible anyway.
Approach: Developers choose when to orchestrate Why Rejected: Inconsistent patterns. Developers often misjudge complexity. Refactoring required when simple becomes complex.
Approach: Move to orchestration incrementally Why Rejected: Why wait? The infrastructure is operational. The performance is proven. The benefits are immediate.
Claim: “0ms orchestration overhead” Confidence Level: MEDIUM (requires production validation) Evidence Type: Development environment measurements Verification Method:
More Accurate Statement: “Orchestration overhead is sub-millisecond and unmeasurable in our current testing environment, representing <0.1% of total operation time”
Production Verification Plan:
The “whole hog” decision represents a philosophical shift: we’re not optimizing for the simple case, we’re optimizing for architectural coherence. When every operation uses the same pattern, the system becomes predictable, maintainable, and evolvable.
The concern that orchestration is “too heavy” for simple tasks was definitively disproven by PM-033d’s 0ms overhead achievement. The concern about token costs was eliminated by Chain-of-Draft’s 92% reduction. There are no remaining technical objections.
This decision embraces the reality that in modern AI systems, coordination is cheap, consistency is valuable, and architectural simplicity pays compound dividends.
Status corrected Accepted → Superseded. “If it can be orchestrated, it should be orchestrated”
was never formally rescinded, but: ADR-043 (Nov 2025) quietly reinterpreted orchestration as an
application-layer workflow pattern (a partial, unlabeled supersession — labeled now), and
services/orchestration/ was deleted 2026-07-18 as unreachable/fabricating code. The premises
(0ms coordination, Chain-of-Draft economics, 7626x acceleration) were never validated per
ADR-015’s own protocol. Field evidence gathered by the Architectural Review 2026 (comparables Leg
C1) independently identifies orchestration-everywhere as the category’s signature overbuild.