Architecture signals & invalidated assumptions

Scope:

Architecture signals (emergent)

Pattern: Pipeline-based orchestration with explicit execution steps improves architectural clarity compared to large orchestrator classes.
Pattern: Typed execution context (PipelineContext) stabilises multi-step pipelines and enables safe parallel agent execution.
Pattern: Parallel agent execution becomes viable once step isolation and context contracts are enforced.
Pattern: Lightweight deterministic intelligence engines (positioning, trajectory, classification) provide strong product intelligence without requiring complex ML infrastructure.
Anti-pattern: Repository access inside pipeline steps introduces boundary leaks and weakens architecture discipline.
Anti-pattern: Orchestrator-private methods reused by steps create hidden coupling and block modular evolution.
Anti-pattern: Unconditional LLM invocation at agent level leads to uncontrolled cost surfaces.
Heuristic: Architectural audits should be executed continuously during refactors rather than only at milestones.
Heuristic: When a refactor introduces unexpected behaviour, first verify lifecycle execution (pipeline triggers, startup flows) before debugging algorithm logic.
Constraint: Cold-start conditions (no history, partial weeks) must be treated as a first-class architecture concern in behavioural intelligence systems.

Assumption: Snapshot recomputation issues were caused by aggregation logic errors rather than pipeline lifecycle gaps.
Assumption: Parallel agent execution would require significant architectural redesign rather than pipeline isolation.
Assumption: Initial pipeline refactor fully removed step–repository coupling.
Assumption: Trend-based intelligence signals could be safely computed without explicit cold-start handling.

High confidence: pipeline orchestration pattern, typed context contracts, step isolation requirements.
Medium confidence: deterministic intelligence engines as the dominant architectural pattern for behavioural coaching systems.
Tentative: optimal strategy for LLM governance and centralized policy enforcement across agents.