Harness, Part 5: Guardrails
Limits make the agent fit for daily use: confirmation before risky actions, an iteration cap, and a list of hard nos.
What this is about
Since Part 4, your agent can act. That is the moment a toy becomes a tool - and a tool becomes a risk. An agent without limits is a script with full authority. Stage 5 of the staged model therefore draws three kinds of limits, and none of them is distrust - all three are operational safety:
Ask - before risky actions, the agent stops and gets your confirmation. Stop - an iteration cap prevents a lost loop from quietly producing costs. Never - a list of hard nos, anchored twice: as words in the identity and, where checkable, as a hard rule in code.
Which actions are risky, you already answered in Part 4 - the note is waiting in your docs. Today it becomes mechanics.
With this part, Block I is complete: LLM call, history, identity, tools, limits. Your harness then is no longer an experiment, but an agent you can trust with daily work.
How the build works
As before: copy the prompt, hand it to your coding agent, answer its questions. The prompt reads harness-doku.html - including your uneasiness note from Part 4.
This section belongs to the Member tier. You get access to full code, deeper explanations, and rabbit-hole material.
See tiers →It works when
You deliberately trigger a risky action - and your agent asks first instead of acting. On your no, it aborts cleanly and explains itself. That completes Block I: your harness can speak, remember, show character, act - and knows where to stop. From here on, the extensions begin: Part 6: Memory gives it a memory that survives the restart.