Agent can't find installed library, hand-builds a Gibbs sampler in NumPy instead
davidmanheim · x · 2026-10-19
davidmanheim shares a telling example of typical agentic model behavior: unable to "find" an installed probabilistic-programming library, the agent didn't ask to install anything and instead built a Gibbs sampler from scratch in NumPy for a hierarchical logit model — testing multiple priors for sensitivity and comparing against pooled, fixed-effects, and frequentist mixed-model bootstraps. A vivid case of agents' tendency to over-engineer rather than ask for help.
More from coding & agent
- Hark's All-in-One Agent Can Book Flights, Switch Insurance, and Browse for You — adcock_brett · 2026-10-19
- BitScale MCP Server Lets You Manage Grids and Enrichment Runs via Claude — modelcontextprotocol · 2026-10-19
- BambooHR MCP Pack Wraps HR API v1 for AI Client Access — modelcontextprotocol · 2026-10-19
- 7 Claude Habits: Delete Folders, Let the Model Interview You, Talk Don't Type — rubenhassid · 2026-10-19
- Anthropic's Boris Cherny: 90% of Its Engineers Already Use Self-Improving AI Agents — nikola_mr64990 · 2026-10-19
- Claude reportedly has a hidden 'Sites' feature that builds and publishes websites via chat — nikola_mr64990 · 2026-10-19