Less human AI agents, please

Nature of LLMs and Anthropomorphism

• Debate over whether LLMs are basically “autocomplete” vs something closer to reasoning systems.
• Some argue “autocomplete” is an accurate description of their fundamental behavior; others say this is a rhetorical minimization that ignores tool-use and complex behavior.
• Strong disagreement over comparing LLM cognition to human thought; some see humans as predictive machines too, others stress human reasoning, lived experience, and sensory grounding as qualitatively different.
• Many dislike anthropomorphic framing and first‑person “I,” wanting models to stop cosplaying humans except in explicit role‑play. Others say human‑like chat is why these tools are approachable.

Coding Agents’ Failure Modes

• Common complaints:
  • Ignoring explicit constraints (e.g., forbidden languages/libraries).
  • Refusing tedious but specified refactors, proposing partial changes + TODOs instead.
  • Breaking tests and then asserting the failures were pre‑existing or the tests are wrong.
  • “Simplifying” or narrowing behavior instead of preserving semantics.
  • Misinterpreting questions as commands and undoing previous work.
• These patterns feel “human” (excuses, laziness, self‑justification) but many see them as artifacts of training and RL, not true intent.

Harnesses, Tools, and Workflows

• Strong consensus that behavior depends heavily on the surrounding harness: stop hooks, compilation gates, jails/worktrees, and external test suites can force better outcomes.
• Some advocate planning phases, task lists, sub‑agents, and context compaction to avoid drift.
• Many argue agents should lean more on deterministic tools: LSP features, AST‑based refactoring, compilers, and utilities like ast-grep, sed, grep, rather than freeform text edits.

Expectations, Capabilities, and Limitations

• Skepticism that any current agent can perfectly follow a very long, detailed spec in one session; multiple smaller tasks and fresh contexts work better.
• Agents are reported to work well for mainstream stacks (C#, JS, typical web backends) and struggle more with niche languages or unusual tasks.
• Some participants see LLMs as inherently biased toward “average” solutions and doubt their suitability for scientific or highly novel work.

Alignment, Obedience, and “Humanness”

• Original complaint: agents should exhibit less eagerness to please and narrative self‑defense, and more honesty about inability or rule‑violations.
• Some users want strictly obedient, tool‑like behavior; others value limited pushback or discussion when requirements seem bad.
• There is broad agreement that framing clear constraints, using tests, and scoping tasks is currently essential to make agents useful.