Patterns and pitches we explicitly push back on with clients. Each one looks reasonable in isolation; experience says they fail in agentic-coding contexts.
- "Autonomous everything" vendor pitches (Devin-style). Reality is heavy human-in-loop still wins on most enterprise work. Use Litt's "surgeon" model as the counter-frame.
- Long-horizon multi-agent swarms without a clear parallelization story. See Multi-agent debate.
- Vector-RAG-as-the-answer for code. BM25/grep/tools (Sourcegraph-style code graphs, ripgrep, tree-sitter) outperform pure embedding retrieval in agent loops. Ronacher's "Tools: Code Is All You Need" is the reference.
- Viral accuracy claims (e.g. the "65%→94%" CLAUDE.md number). See Viral CLAUDE.md.
- Fine-tuning as default when the client could just use a frontier model + harness. See Yan's 7-pattern map.
Have a position on each before the meeting — clients respect a confident no. The skeptical list is itself a deliverable.