Human-held prompts

A person writes and edits the agent's instructions, skills and permissions; the agent doesn't write its own.

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

The agent's setup (its instructions, skills, rules, permissions and reference files) is kept in a harness that a person edits by hand in a workbench. The agent can propose changes, but it can't apply them. Checks on the agent's trace show whether it did what the harness asked, such as reading a given folder.

Our own runs: since 19 November 2026 our harnesses and loops are edited by hand in two workbench apps, and agents no longer touch them (why letting Claude prompt itself fails). Untested beyond our own runs; it is our own prediction that this guards against memory ratcheting and security spazz outs.

Terms on this page

  • Trace: The recorded steps of one run of an AI agent, in order.

Is a

  • Guard: Something that stops a kind of failure before it reaches the end of an unattended run.

Guards

  • Memory ratcheting: Each correction an agent writes for itself stays, so its behaviour only ever tightens.
  • Ossification: The agent turns a choice that suits today into structure that is hard to change later.
  • Security spazz out: The agent guards against the people it serves, beyond what was asked, and the guards outlive the reason for them.