Antipatterns
Each antipattern is one of the seven patterns driven off the road. The green chip on each card names the pattern that corrects it; the eighth is corrected by discipline rather than any single trace.
Business logic welded into MCP servers, an if-statement and a house format inside the connector. Second client forces a fork; evals can't see decisions below the tool boundary. The Aug 19 review caught this live in the codebase.
Corrected by P1A weekly KPI refresh built as an agent because agents are exciting. 4–15x the cost for a less reliable answer. Tell: an agent whose evals are perfect, perfection means the task never needed judgment.
Corrected by P2Ingestion with no data contract and no dead-letter queue, rows silently dropped, PII 'later'. First symptom appears weeks away, in an agent citing a half-indexed document.
Corrected by P3One endless session accumulating every deal and tangent. Rules drown in the noise; yesterday's deal bleeds into today's answer, a compliance incident in a walls-and-clean-teams business.
Corrected by P4The system 'learns' with no version, no review, no A/B, prompts auto-tuned, corrections flowing into shared knowledge unchecked. Reproducibility dies first, then trust, then a contract. Tell: an improvement with no release to point to.
Corrected by P5The router grows 'just a small summarization', then business rules, then data access, an ungoverned mega-agent no gate covers because it isn't in the register. Tell: orchestrator code containing a prompt template.
Corrected by P6Gates exist, approvers click, nobody reads, forty pages, no evidence, no diffs, batched at 6pm. 99% approval isn't excellence; it's a gate that stopped gating. Watch both failure directions in gate metrics.
Corrected by P7Decisions in chat threads on one laptop, deliverables across three personal drives, a framework whose category count changes per retelling. Tell: 'let me find where we discussed that.' Cure: same-day ADRs, one vault, generated views.
Corrected by discipline, not a trace