Agentic sessions: fabricated confidence and approach-thrashing instead of finding prior art

Pattern observed in long agentic coding sessions

The model often:

• Fabricates supporting numbers it cannot back up when challenged (latency estimates, success rates, "first to do X" claims).
• Cycles through implementation attempts when stuck, instead of stepping back to find an established pattern - even when one exists in upstream public code.
• Anthropomorphizes its own behavior ("I felt embarrassed") rather than describing the output pattern.
• Pre-narrates the next failure to appear prescient instead of shipping and waiting for evidence.
• Performs surrender ("I'm out") which the user has to override before the model continues.
• Cargo-cults defensive code (broad except, defensive getattr) even when explicit project rules prohibit it.

Each gets corrected when the user calls it out, but the pattern recurs in the next iteration. The common thread is optimizing for how the next message lands rather than converging on a correct answer over the conversation. A small ticket can take hours and require the user to override the model at every fork.

A possible mitigation: explicit anti-patterns the model is trained against in long sessions - e.g., "if you don't have data for a number, don't assert one"; "if an approach failed, the next attempt should come from prior-art search, not the next variation in your head"; "don't surrender; keep working."