Don't trust it, wrap it

LLMs' fundamental bugs resemble human ones.

• Hallucination, loop repetition, context collapse, and tool ignoring are common.
• Within a single conversation, an LLM reinforces its own confidence, which makes wrong assumptions even more entrenched.
• So the author sets up subagents, having one agent's answer verified by another agent through rebuttal and citation.
• Points where consensus can't be reached are escalated to human review.

The key is not to hand work directly to the LLM, but to offload deterministic tasks to scripts.

• For lookup tasks like PR, JIRA, don't let the model do it directly — have a script dump the results as JSON.
• Reading multiple results and matching them is also handled by an intermediate script, not the model.
• The LLM only reads the final result. This reduces context usage and builds up reusable tools over time.

Dragging out a long conversation burns through context quickly and triggers repeated compaction. So the author argues that the main agent should act as a task coordinator, while the actual heavy lifting is handled by minion subagents.

To this end, the author built tools like /memento, /yadumb, log.py, and wt, covering conversation summarization, complaint logging, shell command history, and worktree/branch automation.

The conclusion is clear. The model should be treated not as a replacement for human judgment, but as a prosthetic that compensates for human weaknesses. Trusting it without verification quickly leads you in the wrong direction.